Add background memory evaluation, off unless configured (#248)

Ships the real, local, testable part of the memory-evaluation plan
(docs/plans/03-memory-evaluation.md): Maven reads back her own recent
memory on a slow ticker, asks the resident model what it notices, and
records the confident answers as notes.

internal/memeval — not internal/memory/eval.go as the plan says, because
internal/store imports internal/memory for the vector backend and an
evaluator has to read store.Fact/Note/Nudge, which would close the
cycle. Evaluate() gathers RecentFacts/RecentNotes/RecentNudges, prompts
under a GBNF grammar bounded to three {observation, confidence,
suggested_action} objects, drops anything under min_confidence,
deduplicates against what earlier runs wrote, and writes the rest as
notes with source infer:memory-eval. /dash already renders notes with
their source, so the output is visible with no UI change.

cmd/mavend/memoryeval.go drives it on its own goroutine and ticker, not
on the 60s tick: an evaluation is a multi-second round-trip on the same
llama-server that answers voice turns, and it runs hourly at most. The
memory_eval config block is absent by default and absence means the
goroutine does not exist. No llama-server phraser also means no loop —
there is no template fallback, because a "memory evaluation" assembled
from templates is a fixed sentence pretending to be an observation.

What it deliberately cannot do, since this is the feature most likely to
turn Maven into a nag:

  - It cannot speak. No dispatcher reference, no channel, no nudge. An
    observation is a thought she wrote down and he reads on /dash.
    Announcing them is a separate decision with its own opt-in.
  - It cannot act. suggested_action is recorded as text and interpreted
    by nobody — no reminder, routine or fact is created from it.
  - It says nothing about an empty store: no memory means no LLM call,
    so there are no observations invented out of two facts.
  - Its own notes are excluded from the next evaluation's input, and are
    written with a nil embedding so they stay out of the recall pool.

The plan's remaining items (dispatching observations, an /eval IPC
method and trace view, RecentEvents) and the fact that output quality is
entirely unmeasured are written up at the bottom of the plan doc.

cmd/mavend/main.go

@@ -169,6 +169,7 @@ func run(args []string) error {
 		coreAPI    ipc.CoreAPI
 		eco        *ecosystemWiring
 		factWorker *factEnrichmentWorker
+		evalWorker *memoryEvalWorker // nil ⇒ memory evaluation off (the default)
 	)
 
 	if !locked {
@@ -263,6 +264,7 @@ func run(args []string) error {
 		autotuneInterval := time.Duration(cfg.AutotuneInterval)
 		tl = newTickLoop(st, gatherer, dispatcher, phr, rules, tickInterval, repeatInterval, autotuneInterval, cfg.Digest, routinesFromConfig(cfg.Routines), config.MorningRoutinesFromConfig(cfg.MorningRoutines), cfg.PatternProposals)
 		factWorker = newFactEnrichmentWorker(st, eco, time.Duration(cfg.FactEnrichmentInterval))
+		evalWorker = newMemoryEvalWorker(st, phr, cfg)
 
 		coreAPI = &daemonAPI{
 			CoreAPI:          ipc.NewStoreAPI(st),
@@ -440,6 +442,7 @@ func run(args []string) error {
 			autotuneInterval := time.Duration(cfg.AutotuneInterval)
 			tl = newTickLoop(st, gatherer, dispatcher, phr, rules, tickInterval, repeatInterval, autotuneInterval, cfg.Digest, routinesFromConfig(cfg.Routines), config.MorningRoutinesFromConfig(cfg.MorningRoutines), cfg.PatternProposals)
 			factWorker = newFactEnrichmentWorker(st, eco, time.Duration(cfg.FactEnrichmentInterval))
+			evalWorker = newMemoryEvalWorker(st, phr, cfg)
 
 			// Swap the CoreAPI from the locked placeholder to the real store adapter.
 			newAPI := &daemonAPI{
@@ -476,6 +479,13 @@ func run(args []string) error {
 				factWorker.run(ctx)
 			}()
 
+			// Start background memory evaluation (nil unless configured).
+			if evalWorker != nil {
+				go func() {
+					evalWorker.run(ctx)
+				}()
+			}
+
 			dl.unlock()
 			log.Printf("mavend: unlocked via passkey assertion")
 			return nil
@@ -514,6 +524,13 @@ func run(args []string) error {
 			defer wg.Done()
 			factWorker.run(ctx)
 		}()
+		if evalWorker != nil {
+			wg.Add(1)
+			go func() {
+				defer wg.Done()
+				evalWorker.run(ctx)
+			}()
+		}
 	}
 
 	<-ctx.Done()

cmd/mavend/memoryeval.go

@@ -0,0 +1,88 @@
+// mavend/memoryeval.go — the driver for background memory evaluation
+// (Vikunja #248). The evaluator itself is pure-ish and lives in
+// internal/memeval; this is the one impure part: a ticker, the store, and the
+// resident model's base URL.
+//
+// It is its own goroutine and NOT a step on the main tick, deliberately. The
+// tick runs every 60s and has a delivery deadline behind it; an evaluation is
+// a multi-second LLM round-trip on the same llama-server that answers voice
+// turns, and it happens hourly at most. Bolting it onto the tick would make
+// every hour's tick the slow one for no benefit.
+package main
+
+import (
+	"context"
+	"log"
+	"time"
+
+	"github.com/kami/maven/internal/config"
+	"github.com/kami/maven/internal/llm"
+	"github.com/kami/maven/internal/memeval"
+	"github.com/kami/maven/internal/phraser"
+	"github.com/kami/maven/internal/store"
+)
+
+// memoryEvalWorker — ticker + evaluator.
+type memoryEvalWorker struct {
+	eval     *memeval.Evaluator
+	interval time.Duration
+}
+
+// newMemoryEvalWorker wires the evaluation loop, or returns nil when it should
+// not run at all. nil is the normal case and every caller must handle it:
+//
+//   - no memory_eval config block ⇒ off (a capability is off unless configured);
+//   - no LLM phraser ⇒ nothing to evaluate with. There is no template fallback
+//     here on purpose: a "memory evaluation" assembled from string templates
+//     would be a fixed sentence pretending to be an observation.
+func newMemoryEvalWorker(st *store.Store, phr phraser.Phraser, cfg *config.Config) *memoryEvalWorker {
+	if cfg.MemoryEval == nil {
+		return nil
+	}
+	lp, ok := phr.(*phraser.LLMPhraser)
+	if !ok {
+		log.Printf("memory eval: configured but no llama-server phraser — evaluation disabled")
+		return nil
+	}
+	interval := time.Duration(cfg.MemoryEval.Interval)
+	if interval <= 0 {
+		interval = config.DefaultMemoryEvalInterval
+	}
+	// A generous per-request timeout: this is a long prompt to a Thinking model
+	// and nobody is waiting on the answer.
+	client := llm.New(lp.BaseURL(), 5*time.Minute)
+	ev := memeval.NewEvaluator(st, st, client, memeval.Config{
+		MaxItems:      cfg.MemoryEval.MaxItems,
+		MinConfidence: cfg.MemoryEval.MinConfidence,
+		ContextBlock:  contextBlockFn(cfg, time.Now),
+	})
+	log.Printf("memory eval: enabled, every %s", interval)
+	return &memoryEvalWorker{eval: ev, interval: interval}
+}
+
+// run evaluates every interval until ctx is canceled.
+//
+// The first evaluation waits a full interval rather than firing at startup, the
+// opposite of the tick loop's cold-start behaviour. A tick that fires late is a
+// nudge that arrives late; an evaluation that fires late is nothing at all, and
+// the alternative is a heavy LLM call competing with startup — including with
+// the first voice turn after a restart.
+func (w *memoryEvalWorker) run(ctx context.Context) {
+	ticker := time.NewTicker(w.interval)
+	defer ticker.Stop()
+	for {
+		select {
+		case <-ctx.Done():
+			return
+		case now := <-ticker.C:
+			obs, err := w.eval.Evaluate(ctx, now)
+			if err != nil {
+				log.Printf("memory eval: %v", err)
+				continue
+			}
+			for _, o := range obs {
+				log.Printf("memory eval: noted (%.2f, %s): %s", o.Conf, o.Action, o.Text)
+			}
+		}
+	}
+}

docs/plans/03-memory-evaluation.md

@@ -25,3 +25,46 @@
 6. Add `/eval` API method to `ipc.CoreAPI` (or reuse `Chat` with system context) so mavweb can show evaluation history
 7. Add `memory_eval` block to `deploy/mavend.json`
 8. Test with synthetic store state — verify observations match expected patterns
+
+---
+
+## Status 2026-08-01 — foundation shipped (Vikunja #248)
+
+**Shipped:** `internal/memeval` (not `internal/memory/eval.go` — `internal/store`
+imports `internal/memory` for the vector backend, so an evaluator that reads
+`store.Fact` there would close an import cycle). `Evaluator.Evaluate` reads
+`RecentFacts` / `RecentNotes` / `RecentNudges`, prompts the resident model under
+a GBNF grammar for at most three `{observation, confidence, suggested_action}`
+objects, drops anything under `min_confidence`, deduplicates against what earlier
+evaluations wrote, and records the rest as notes with source `infer:memory-eval`.
+Driver: `cmd/mavend/memoryeval.go`, its own goroutine on its own ticker. Config:
+the `memory_eval` block — **absent ⇒ the loop does not run**. Visibility: `/dash`
+already renders notes with their source, so evaluation output is visible with no
+UI change.
+
+**Deliberately not shipped — this is policy, not an unfinished edge:**
+
+- *Dispatching observations as care nudges (plan step 4).* An hourly LLM loop
+  with permission to speak is a machine for generating interruptions, and the
+  content is model-generated text about his own life. The evaluator has no
+  dispatcher reference at all, so it cannot reach a channel by accident. Wiring
+  it to `delivery.Dispatcher` is a separate decision with its own opt-in.
+- *Acting on `suggested_action`.* It is recorded inside the note text and
+  interpreted by nobody. No reminder, routine or fact is created.
+- *Writing observation embeddings.* Notes are written with a nil embedding, so
+  they stay out of the RAG recall pool. Feeding generated text back into the pool
+  it came from is how a small model starts citing its own guesses as evidence.
+
+**Deferred, wants a decision or another capability:**
+
+- *Plan step 6, the `/eval` IPC method and an evaluation-history view.* `/dash`
+  covers reading the output; a dedicated trace surface is worth building once
+  there is real output to look at, and it should probably show the prompt too.
+- *`RecentEvents`.* The plan lists it; the evaluator reads facts, notes and
+  nudges. Detected action/object events already drive pattern proposals (#43), and
+  duplicating them here would mostly re-derive that.
+- *Output quality is unmeasured.* There is no fixture for "did she notice
+  something true". The tests cover the machinery — empty store, confidence floor,
+  dedupe, own-notes exclusion, error handling — not the observations. Until
+  someone reads a week of real output on `/dash`, treat the wording and the
+  `min_confidence` default as unvalidated.

internal/config/config.go

@@ -146,6 +146,10 @@ type Config struct {
 	// PatternProposalConfig.
 	PatternProposals *PatternProposalConfig `json:"pattern_proposals,omitempty"`
 
+	// MemoryEval — background memory evaluation (internal/memeval). nil /
+	// absent ⇒ no evaluation loop at all. See MemoryEvalConfig.
+	MemoryEval *MemoryEvalConfig `json:"memory_eval,omitempty"`
+
 	// Praxis — the ecosystem attention-state service. When configured, maven
 	// calls the Praxis HTTP tools API for attention listing and item lifecycle.
 	// Maven never touches Praxis's database directly (ecosystem invariant: no
@@ -392,6 +396,28 @@ func (p *PatternProposalConfig) AnnounceProposals() bool {
 	return p != nil && p.Notify
 }
 
+// MemoryEvalConfig — the background memory-evaluation loop (Vikunja #248).
+// Absent ⇒ off, like every other capability that costs something the owner did
+// not ask for. Each evaluation is a full LLM round-trip on the one resident
+// model, which is the same model answering him; running it hourly by default
+// would put a multi-second stall in front of an occasional voice turn for a
+// feature he may not want.
+//
+// The loop only ever writes notes (source infer:memory-eval, visible on
+// /dash). It cannot speak — see internal/memeval.
+type MemoryEvalConfig struct {
+	// Interval — how often to evaluate. 0 ⇒ DefaultMemoryEvalInterval.
+	Interval Duration `json:"interval,omitempty"`
+
+	// MaxItems — recent facts / notes / nudges fed into one evaluation.
+	// 0 ⇒ memeval.DefaultMaxItems.
+	MaxItems int `json:"max_items,omitempty"`
+
+	// MinConfidence — observations the model scores below this are dropped.
+	// 0 ⇒ memeval.DefaultMinConfidence.
+	MinConfidence float64 `json:"min_confidence,omitempty"`
+}
+
 // PhraserConfig — the LLM-backed phraser seam. The daemon spawns llama-server
 // as a managed subprocess and sends chat-completion requests to phrase nudge
 // and reminder messages. nil ⇒ the template-based Stub is used instead.
@@ -493,6 +519,10 @@ const (
 	// most. A proposal is never urgent; if two patterns surface in the same
 	// hour, the second one waits, and the /routines page has it either way.
 	DefaultProposalCooldown = 24 * time.Hour
+
+	// DefaultMemoryEvalInterval — the plan's cadence (1h) for the memory
+	// evaluation loop, applied only when the block is present at all.
+	DefaultMemoryEvalInterval = time.Hour
 )
 
 // Load reads the JSON config at path and applies defaults. A missing file is
@@ -571,6 +601,12 @@ func (c *Config) applyDefaults() {
 		c.PatternProposals.Cooldown = Duration(DefaultProposalCooldown)
 	}
 
+	// Same rule: absent stays nil (⇒ no evaluation loop), present gets defaults
+	// so `{}` is a valid "on with the plan's cadence".
+	if c.MemoryEval != nil && c.MemoryEval.Interval <= 0 {
+		c.MemoryEval.Interval = Duration(DefaultMemoryEvalInterval)
+	}
+
 	if c.Voice != nil {
 		if c.Voice.RouterThreshold <= 0 {
 			c.Voice.RouterThreshold = DefaultRouterThreshold

internal/config/config_test.go

@@ -243,3 +243,53 @@ func TestDurationRoundTrip(t *testing.T) {
 		t.Errorf("round-trip = %v, want %v", d2, d)
 	}
 }
+
+// Both new opt-in capabilities follow the same rule: absent block ⇒ nil ⇒ the
+// behaviour does not exist. Presence is the enable act, so a bare `{}` block is
+// valid and gets the defaults filled in.
+func TestOptInBlocksAbsentStayNil(t *testing.T) {
+	c, err := Load(writeConfig(t, `{}`))
+	if err != nil {
+		t.Fatalf("Load: %v", err)
+	}
+	if c.PatternProposals != nil {
+		t.Errorf("pattern_proposals absent but got %+v", c.PatternProposals)
+	}
+	if c.PatternProposals.AnnounceProposals() {
+		t.Error("AnnounceProposals() true with no config block")
+	}
+	if c.MemoryEval != nil {
+		t.Errorf("memory_eval absent but got %+v", c.MemoryEval)
+	}
+}
+
+func TestOptInBlocksGetDefaultsWhenPresent(t *testing.T) {
+	c, err := Load(writeConfig(t, `{"pattern_proposals":{"notify":true},"memory_eval":{}}`))
+	if err != nil {
+		t.Fatalf("Load: %v", err)
+	}
+	if !c.PatternProposals.AnnounceProposals() {
+		t.Error("notify:true did not enable announcements")
+	}
+	if time.Duration(c.PatternProposals.Cooldown) != DefaultProposalCooldown {
+		t.Errorf("proposal cooldown = %v, want %v", c.PatternProposals.Cooldown, DefaultProposalCooldown)
+	}
+	if time.Duration(c.MemoryEval.Interval) != DefaultMemoryEvalInterval {
+		t.Errorf("memory eval interval = %v, want %v", c.MemoryEval.Interval, DefaultMemoryEvalInterval)
+	}
+}
+
+// Notify is off even when the block exists — the block is where you tune it,
+// notify:true is the act that lets her speak.
+func TestPatternProposalNotifyDefaultsOff(t *testing.T) {
+	c, err := Load(writeConfig(t, `{"pattern_proposals":{"cooldown":"6h"}}`))
+	if err != nil {
+		t.Fatalf("Load: %v", err)
+	}
+	if c.PatternProposals.AnnounceProposals() {
+		t.Error("notify defaulted to on")
+	}
+	if time.Duration(c.PatternProposals.Cooldown) != 6*time.Hour {
+		t.Errorf("cooldown = %v, want 6h", c.PatternProposals.Cooldown)
+	}
+}

internal/memeval/eval.go

@@ -0,0 +1,357 @@
+// Package memeval is background memory evaluation (Vikunja #248,
+// docs/plans/03-memory-evaluation.md).
+//
+// It lives beside internal/memory rather than inside it because
+// internal/store imports internal/memory for the vector-store backend, and an
+// evaluator has to read store.Fact / store.Note / store.Nudge — putting it in
+// internal/memory would close that import cycle.
+//
+// Every so often Maven reads back her own recent memory — facts, notes, the
+// nudges she sent — and asks the resident model what it notices: a habit that
+// stopped, a gap, something worth saying later. What comes back is written as
+// notes with source EvalNoteSource and nothing else happens. That restraint is
+// the design, not an unfinished edge:
+//
+//   - She does not speak here. There is no dispatcher, no channel, no nudge.
+//     An observation is a thought she wrote down; he reads it on /dash when he
+//     wants to. "Not a nag, not autonomous" (CLAUDE.md) is easy to violate with
+//     exactly this feature — an hourly loop with an LLM in it and permission to
+//     talk is a machine for generating interruptions — so the loop has no way
+//     to reach him at all. Turning observations into nudges is a separate
+//     decision with a separate opt-in, and it is deliberately NOT in this file.
+//   - She does not act. No reminder is created, no routine proposed, no fact
+//     written. The model's suggested_action is recorded as text inside the note
+//     and interpreted by nobody.
+//   - She says nothing about an empty store. No memory ⇒ no LLM call ⇒ no
+//     "observations" invented out of two facts. A 1.7B asked to find a pattern
+//     will always find one; the defence is not asking.
+//
+// Everything the evaluator writes is attributable: source is EvalNoteSource, so
+// an inferred observation can never be mistaken for something he said, and the
+// whole batch is one SQL delete away if the output turns out to be noise.
+package memeval
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+	"sort"
+	"strings"
+	"time"
+
+	"github.com/kami/maven/internal/llm"
+	"github.com/kami/maven/internal/persona"
+	"github.com/kami/maven/internal/store"
+)
+
+// EvalNoteSource — the source stamped on every note the evaluator writes.
+// Same infer:* convention as the rest of the derived facts.
+const EvalNoteSource = "infer:memory-eval"
+
+// DefaultMinConfidence — an observation below this is dropped. The model is
+// asked for its own confidence and small models are badly calibrated, so this
+// is a coarse filter, not a probability: it exists to throw away the guesses
+// the model itself hedged on.
+const DefaultMinConfidence = 0.7
+
+// DefaultMaxItems — how much recent memory goes into one evaluation, per
+// store. 30 facts + 30 notes + 30 nudges is a few thousand tokens of the 4096
+// context the resident Thinking model runs with, which leaves room for its
+// reasoning tokens. Raising this trades reasoning room for history.
+const DefaultMaxItems = 30
+
+// MaxObservations — the model may return at most this many observations per
+// evaluation, enforced by the grammar. A cap here is also a noise cap: an
+// evaluation that "notices" ten things has noticed nothing.
+const MaxObservations = 3
+
+// Observation — one thing the evaluator noticed.
+type Observation struct {
+	Text string  `json:"observation"`
+	Conf float64 `json:"confidence"`
+	// Action — what the model thinks should happen with this. Recorded, never
+	// executed: see the file comment. One of "note", "propose", "notify".
+	Action string `json:"suggested_action"`
+}
+
+// Completer — the llama-server seam, same shape router.Completer uses so the
+// one resident model serves this caller too.
+type Completer interface {
+	Complete(ctx context.Context, r llm.Req) (string, error)
+}
+
+// Reader — the slice of the store an evaluation reads. Narrow on purpose: the
+// evaluator gets recent memory and nothing else. No entity graph, no presence,
+// no config facts.
+type Reader interface {
+	RecentFacts(ctx context.Context, n int) ([]store.Fact, error)
+	RecentNotes(ctx context.Context, n int) ([]store.Note, error)
+	RecentNudges(ctx context.Context, n int) ([]store.Nudge, error)
+}
+
+// NoteWriter — where observations land. Embeddings are passed nil: an
+// observation is written for a human to read on /dash, not to be recalled by
+// similarity. Feeding LLM-generated text back into the RAG pool it was
+// generated from is how a small model starts citing its own guesses as
+// evidence.
+type NoteWriter interface {
+	WriteNote(ctx context.Context, ts time.Time, text string, embedding []float32, source string) (int64, error)
+}
+
+// Config — evaluator tuning. Zero values are replaced by the Default*
+// constants, so the zero Config is the sane one.
+type Config struct {
+	MaxItems      int
+	MinConfidence float64
+
+	// ContextBlock — the shared persona block (internal/persona), re-evaluated
+	// per call so the clock in it is current. Prepended to the system prompt so
+	// observations come out in Maven's voice: feminine self-reference, informal
+	// "ты". nil is allowed; the base prompt still carries the address rules.
+	ContextBlock func() string
+}
+
+// Evaluator reads recent memory and records what the model notices.
+type Evaluator struct {
+	read  Reader
+	write NoteWriter
+	llm   Completer
+	cfg   Config
+}
+
+func NewEvaluator(r Reader, w NoteWriter, c Completer, cfg Config) *Evaluator {
+	if cfg.MaxItems <= 0 {
+		cfg.MaxItems = DefaultMaxItems
+	}
+	if cfg.MinConfidence <= 0 {
+		cfg.MinConfidence = DefaultMinConfidence
+	}
+	return &Evaluator{read: r, write: w, llm: c, cfg: cfg}
+}
+
+// evalGrammar — GBNF pinning the reply to a bounded JSON array of fixed-shape
+// observations. Same reasoning as the router's routeGrammar: the enum and the
+// length bound are what stop a small model from drifting into free text or
+// filling the token budget with one repeated field.
+const evalGrammar = `
+root   ::= "[" ws (obs ("," ws obs){0,2})? ws "]"
+obs    ::= "{" ws "\"observation\"" ws ":" ws text "," ws "\"confidence\"" ws ":" ws conf "," ws "\"suggested_action\"" ws ":" ws act ws "}"
+text   ::= "\"" ([^"\\] | "\\" .){1,200} "\""
+conf   ::= "0" "." [0-9]{1,2} | "1" ("." "0")?
+act    ::= "\"note\"" | "\"propose\"" | "\"notify\""
+ws     ::= [ \t\n]*
+`
+
+// evalSystem — the evaluation prompt. Two things it insists on, both learned
+// from the phraser: state the observation as something she noticed rather than
+// an instruction, and say nothing when there is nothing (the model is given an
+// explicit way to return an empty array, because a model with no exit returns
+// filler).
+const evalSystem = `Ты просматриваешь свою собственную память: недавние факты, заметки и напоминания, которые ты отправляла.
+Найди то, что действительно заметно: привычка, которая прервалась; пробел в записях; повторяющаяся закономерность.
+
+Правила:
+- Отвечай ТОЛЬКО массивом JSON. Каждый элемент: {"observation": "...", "confidence": 0.0-1.0, "suggested_action": "note"|"propose"|"notify"}.
+- observation — короткая фраза по-русски о том, что ты заметила. О себе — в женском роде ("я заметила"). К нему — на "ты".
+- Не выдумывай. Если в памяти нет ничего заметного, верни пустой массив [].
+- Не давай советов и не приказывай. Ты замечаешь, а не требуешь.
+- confidence — насколько ты уверена, что это настоящая закономерность, а не совпадение.
+- Максимум три наблюдения. Лучше одно точное, чем три общих.`
+
+// Evaluate runs one evaluation and returns the observations it recorded.
+//
+// Returns (nil, nil) — not an error — for every ordinary "nothing to say"
+// outcome: an empty store, an empty array from the model, everything below the
+// confidence floor, or every observation already recorded earlier. Only a real
+// read/LLM/write failure is an error, and the caller (a background ticker) logs
+// it and waits for the next interval.
+func (e *Evaluator) Evaluate(ctx context.Context, now time.Time) ([]Observation, error) {
+	snap, err := e.snapshot(ctx)
+	if err != nil {
+		return nil, err
+	}
+	if snap == "" {
+		return nil, nil // nothing recorded ⇒ nothing to notice, and no LLM call
+	}
+
+	raw, err := e.llm.Complete(ctx, llm.Req{
+		System:        persona.Prepend(e.cfg.ContextBlock, evalSystem),
+		User:          snap,
+		Grammar:       evalGrammar,
+		MaxTokens:     512,
+		RepeatPenalty: 1.1,
+	})
+	if err != nil {
+		return nil, fmt.Errorf("memory eval: complete: %w", err)
+	}
+	obs, err := parseObservations(raw)
+	if err != nil {
+		return nil, fmt.Errorf("memory eval: parse %q: %w", truncate(raw, 120), err)
+	}
+
+	// Dedupe against what earlier evaluations already wrote. Without this an
+	// hourly loop over a slowly-changing store writes the same sentence every
+	// hour until /dash is nothing but the evaluator talking to itself.
+	seen, err := e.recordedTexts(ctx)
+	if err != nil {
+		return nil, err
+	}
+
+	var kept []Observation
+	for _, o := range obs {
+		o.Text = strings.TrimSpace(o.Text)
+		if o.Text == "" || o.Conf < e.cfg.MinConfidence {
+			continue
+		}
+		norm := normalizeObservation(o.Text)
+		if seen[norm] {
+			continue
+		}
+		seen[norm] = true
+		if _, err := e.write.WriteNote(ctx, now, formatNote(o), nil, EvalNoteSource); err != nil {
+			return kept, fmt.Errorf("memory eval: write note: %w", err)
+		}
+		kept = append(kept, o)
+	}
+	return kept, nil
+}
+
+// formatNote — the stored text. The suggested action is kept as a visible
+// suffix rather than a column: it is the model's opinion about what to do next,
+// and the only consumer is a human reading /dash.
+func formatNote(o Observation) string {
+	if o.Action == "" {
+		return o.Text
+	}
+	return fmt.Sprintf("%s [%s]", o.Text, o.Action)
+}
+
+// recordedTexts — the normalized text of every observation earlier evaluations
+// wrote, for dedupe. Reads a wider window than MaxItems because the point is to
+// remember saying it, not to summarize it.
+func (e *Evaluator) recordedTexts(ctx context.Context) (map[string]bool, error) {
+	notes, err := e.read.RecentNotes(ctx, 200)
+	if err != nil {
+		return nil, fmt.Errorf("memory eval: recent notes: %w", err)
+	}
+	seen := make(map[string]bool, len(notes))
+	for _, n := range notes {
+		if n.Source != EvalNoteSource {
+			continue
+		}
+		text := n.Text
+		// Strip the "[action]" suffix formatNote appended.
+		if i := strings.LastIndex(text, " ["); i > 0 && strings.HasSuffix(text, "]") {
+			text = text[:i]
+		}
+		seen[normalizeObservation(text)] = true
+	}
+	return seen, nil
+}
+
+// normalizeObservation — dedupe key. Case- and whitespace-insensitive, which
+// catches the realistic repeat (the model re-emitting the same sentence with a
+// different comma) without pretending to do semantic dedupe.
+func normalizeObservation(s string) string {
+	return strings.Join(strings.Fields(strings.ToLower(s)), " ")
+}
+
+// snapshot renders recent memory as the user turn. Returns "" when there is
+// nothing in any store — the caller treats that as "do not ask the model".
+//
+// Notes written by earlier evaluations are excluded. Feeding her own
+// observations back in is how "я заметила, что ты не записывал еду" becomes
+// evidence for noticing it again, three evaluations deep.
+func (e *Evaluator) snapshot(ctx context.Context) (string, error) {
+	n := e.cfg.MaxItems
+	facts, err := e.read.RecentFacts(ctx, n)
+	if err != nil {
+		return "", fmt.Errorf("memory eval: recent facts: %w", err)
+	}
+	notes, err := e.read.RecentNotes(ctx, n)
+	if err != nil {
+		return "", fmt.Errorf("memory eval: recent notes: %w", err)
+	}
+	nudges, err := e.read.RecentNudges(ctx, n)
+	if err != nil {
+		return "", fmt.Errorf("memory eval: recent nudges: %w", err)
+	}
+
+	var b strings.Builder
+	wrote := false
+	if len(facts) > 0 {
+		b.WriteString("Факты:\n")
+		for _, f := range facts {
+			fmt.Fprintf(&b, "- %s %s=%s (%s)\n", f.Ts.Format("2006-01-02 15:04"), f.Key, truncate(f.Value, 80), f.Source)
+			wrote = true
+		}
+	}
+	own := 0
+	var noteLines []string
+	for _, nt := range notes {
+		if nt.Source == EvalNoteSource {
+			own++
+			continue
+		}
+		noteLines = append(noteLines, fmt.Sprintf("- %s %s\n", nt.Ts.Format("2006-01-02 15:04"), truncate(nt.Text, 160)))
+	}
+	if len(noteLines) > 0 {
+		b.WriteString("\nЗаметки:\n")
+		for _, l := range noteLines {
+			b.WriteString(l)
+			wrote = true
+		}
+	}
+	if len(nudges) > 0 {
+		b.WriteString("\nНапоминания, которые ты отправляла:\n")
+		for _, nd := range nudges {
+			outcome := nd.Outcome
+			if outcome == "" {
+				outcome = "?"
+			}
+			fmt.Fprintf(&b, "- %s %s → %s (%s)\n", nd.Ts.Format("2006-01-02 15:04"), nd.Rule, outcome, nd.Channel)
+			wrote = true
+		}
+	}
+	if !wrote {
+		// Only her own past observations, or nothing at all. Either way there is
+		// no new memory to evaluate.
+		return "", nil
+	}
+	b.WriteString("\nЧто ты замечаешь?")
+	return b.String(), nil
+}
+
+// parseObservations reads the model's array. Tolerates the leading/trailing
+// prose a Thinking model sometimes emits around JSON by taking the outermost
+// bracketed span, the same tolerance the router's parser has.
+func parseObservations(raw string) ([]Observation, error) {
+	s := strings.TrimSpace(raw)
+	if i := strings.Index(s, "["); i >= 0 {
+		if j := strings.LastIndex(s, "]"); j > i {
+			s = s[i : j+1]
+		}
+	}
+	if s == "" {
+		return nil, nil
+	}
+	var obs []Observation
+	if err := json.Unmarshal([]byte(s), &obs); err != nil {
+		return nil, err
+	}
+	if len(obs) > MaxObservations {
+		// The grammar bounds this; a grammar-less server or a future prompt
+		// change must not be able to flood /dash.
+		sort.SliceStable(obs, func(i, j int) bool { return obs[i].Conf > obs[j].Conf })
+		obs = obs[:MaxObservations]
+	}
+	return obs, nil
+}
+
+func truncate(s string, n int) string {
+	r := []rune(s)
+	if len(r) <= n {
+		return s
+	}
+	return string(r[:n]) + "…"
+}

internal/memeval/eval_test.go

@@ -0,0 +1,271 @@
+package memeval
+
+import (
+	"context"
+	"database/sql"
+	"errors"
+	"path/filepath"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/kami/maven/internal/llm"
+	"github.com/kami/maven/internal/store"
+)
+
+// fakeLLM — canned replies, one per call, and a record of what it was asked.
+type fakeLLM struct {
+	replies []string
+	calls   []llm.Req
+	err     error
+}
+
+func (f *fakeLLM) Complete(_ context.Context, r llm.Req) (string, error) {
+	f.calls = append(f.calls, r)
+	if f.err != nil {
+		return "", f.err
+	}
+	if len(f.replies) == 0 {
+		return "[]", nil
+	}
+	out := f.replies[0]
+	f.replies = f.replies[1:]
+	return out, nil
+}
+
+func newTestStore(t *testing.T) *store.Store {
+	t.Helper()
+	st, err := store.Open(context.Background(), filepath.Join(t.TempDir(), "memeval_test.db"))
+	if err != nil {
+		t.Fatalf("store.Open: %v", err)
+	}
+	t.Cleanup(func() { _ = st.Close() })
+	return st
+}
+
+func refNow() time.Time { return time.Date(2026, 8, 1, 9, 0, 0, 0, time.UTC) }
+
+// seedMemory writes a little of everything the evaluator reads.
+func seedMemory(t *testing.T, st *store.Store, ctx context.Context, now time.Time) {
+	t.Helper()
+	for i := 0; i < 3; i++ {
+		ts := now.Add(-time.Duration(i+1) * 24 * time.Hour)
+		if _, err := st.WriteFact(ctx, ts, store.KindSelf, "water_ml", "500", "tap:desk", 1.0, sql.NullInt64{}); err != nil {
+			t.Fatalf("write fact: %v", err)
+		}
+	}
+	if _, err := st.WriteNote(ctx, now.Add(-2*time.Hour), "купить корм для кота", nil, "tap:voice"); err != nil {
+		t.Fatalf("write note: %v", err)
+	}
+	if _, err := st.RecordNudge(ctx, "water", "voice", "пора выпить воды", now.Add(-time.Hour)); err != nil {
+		t.Fatalf("record nudge: %v", err)
+	}
+}
+
+// TestEvaluateEmptyStoreAsksNothing — the "shuts up when uncertain" floor. An
+// empty store must not even reach the model: a small model asked to find a
+// pattern in nothing will invent one.
+func TestEvaluateEmptyStoreAsksNothing(t *testing.T) {
+	st := newTestStore(t)
+	ctx := context.Background()
+	f := &fakeLLM{}
+	ev := NewEvaluator(st, st, f, Config{})
+
+	obs, err := ev.Evaluate(ctx, refNow())
+	if err != nil {
+		t.Fatalf("Evaluate: %v", err)
+	}
+	if len(obs) != 0 {
+		t.Fatalf("observations on an empty store = %d, want 0", len(obs))
+	}
+	if len(f.calls) != 0 {
+		t.Fatalf("LLM called %d times on an empty store, want 0", len(f.calls))
+	}
+}
+
+// TestEvaluateWritesHighConfidenceObservations — the happy path. Confident
+// observations are written as notes stamped infer:memory-eval, and the low
+// ones are dropped.
+func TestEvaluateWritesHighConfidenceObservations(t *testing.T) {
+	st := newTestStore(t)
+	ctx := context.Background()
+	now := refNow()
+	seedMemory(t, st, ctx, now)
+
+	f := &fakeLLM{replies: []string{`[
+		{"observation":"ты три дня не записывал еду","confidence":0.9,"suggested_action":"notify"},
+		{"observation":"может быть, ты стал меньше пить воды","confidence":0.3,"suggested_action":"note"}
+	]`}}
+	ev := NewEvaluator(st, st, f, Config{})
+
+	obs, err := ev.Evaluate(ctx, now)
+	if err != nil {
+		t.Fatalf("Evaluate: %v", err)
+	}
+	if len(obs) != 1 {
+		t.Fatalf("kept %d observations, want 1 (the 0.3 one is below the floor): %+v", len(obs), obs)
+	}
+	if obs[0].Text != "ты три дня не записывал еду" {
+		t.Errorf("kept the wrong observation: %q", obs[0].Text)
+	}
+
+	notes, err := st.RecentNotes(ctx, 50)
+	if err != nil {
+		t.Fatalf("RecentNotes: %v", err)
+	}
+	var written []store.Note
+	for _, n := range notes {
+		if n.Source == EvalNoteSource {
+			written = append(written, n)
+		}
+	}
+	if len(written) != 1 {
+		t.Fatalf("notes with source %s = %d, want 1", EvalNoteSource, len(written))
+	}
+	if !strings.Contains(written[0].Text, "ты три дня не записывал еду") {
+		t.Errorf("note text = %q", written[0].Text)
+	}
+	if !strings.Contains(written[0].Text, "[notify]") {
+		t.Errorf("note text = %q, want the suggested action recorded", written[0].Text)
+	}
+
+	// The prompt must carry the memory it is evaluating, and must not carry a
+	// grammar-free request.
+	if len(f.calls) != 1 {
+		t.Fatalf("LLM calls = %d, want 1", len(f.calls))
+	}
+	if !strings.Contains(f.calls[0].User, "water_ml") {
+		t.Errorf("prompt does not mention the seeded facts:\n%s", f.calls[0].User)
+	}
+	if f.calls[0].Grammar == "" {
+		t.Error("evaluation ran without a grammar")
+	}
+}
+
+// TestEvaluateDeduplicatesAcrossRuns — the failure mode that would make this
+// feature unusable: an hourly loop over a store that barely changes writing the
+// same sentence every hour until /dash is nothing but the evaluator.
+func TestEvaluateDeduplicatesAcrossRuns(t *testing.T) {
+	st := newTestStore(t)
+	ctx := context.Background()
+	now := refNow()
+	seedMemory(t, st, ctx, now)
+
+	same := `[{"observation":"ты три дня не записывал еду","confidence":0.9,"suggested_action":"note"}]`
+	spaced := `[{"observation":"Ты  три дня   не записывал еду","confidence":0.95,"suggested_action":"note"}]`
+	f := &fakeLLM{replies: []string{same, same, spaced}}
+	ev := NewEvaluator(st, st, f, Config{})
+
+	for i := 0; i < 3; i++ {
+		if _, err := ev.Evaluate(ctx, now.Add(time.Duration(i)*time.Hour)); err != nil {
+			t.Fatalf("Evaluate %d: %v", i, err)
+		}
+	}
+
+	notes, err := st.RecentNotes(ctx, 50)
+	if err != nil {
+		t.Fatalf("RecentNotes: %v", err)
+	}
+	n := 0
+	for _, nt := range notes {
+		if nt.Source == EvalNoteSource {
+			n++
+		}
+	}
+	if n != 1 {
+		t.Fatalf("eval notes after three identical evaluations = %d, want 1", n)
+	}
+}
+
+// TestEvaluateIgnoresOwnNotes — her own observations must not become input.
+// Otherwise "я заметила X" is evidence for noticing X again, three evaluations
+// deep. With nothing but eval notes in the store there is no new memory, so the
+// model is not asked at all.
+func TestEvaluateIgnoresOwnNotes(t *testing.T) {
+	st := newTestStore(t)
+	ctx := context.Background()
+	now := refNow()
+	if _, err := st.WriteNote(ctx, now.Add(-time.Hour), "я заметила, что ты мало пьёшь [note]", nil, EvalNoteSource); err != nil {
+		t.Fatalf("write note: %v", err)
+	}
+
+	f := &fakeLLM{}
+	ev := NewEvaluator(st, st, f, Config{})
+	obs, err := ev.Evaluate(ctx, now)
+	if err != nil {
+		t.Fatalf("Evaluate: %v", err)
+	}
+	if len(obs) != 0 || len(f.calls) != 0 {
+		t.Fatalf("observations=%d llm calls=%d, want 0/0 — own notes are not memory to evaluate", len(obs), len(f.calls))
+	}
+}
+
+// TestEvaluateEmptyArrayIsNotAnError — "nothing to say" is the expected outcome
+// most of the time and must not be logged as a failure.
+func TestEvaluateEmptyArrayIsNotAnError(t *testing.T) {
+	st := newTestStore(t)
+	ctx := context.Background()
+	now := refNow()
+	seedMemory(t, st, ctx, now)
+
+	ev := NewEvaluator(st, st, &fakeLLM{replies: []string{"[]"}}, Config{})
+	obs, err := ev.Evaluate(ctx, now)
+	if err != nil {
+		t.Fatalf("Evaluate: %v", err)
+	}
+	if len(obs) != 0 {
+		t.Fatalf("observations = %d, want 0", len(obs))
+	}
+}
+
+// TestEvaluateLLMErrorIsReported — a broken llama-server is an error the caller
+// logs; it must not silently write anything.
+func TestEvaluateLLMErrorIsReported(t *testing.T) {
+	st := newTestStore(t)
+	ctx := context.Background()
+	now := refNow()
+	seedMemory(t, st, ctx, now)
+
+	ev := NewEvaluator(st, st, &fakeLLM{err: errors.New("connection refused")}, Config{})
+	if _, err := ev.Evaluate(ctx, now); err == nil {
+		t.Fatal("want an error when the model is unreachable")
+	}
+	notes, err := st.RecentNotes(ctx, 50)
+	if err != nil {
+		t.Fatalf("RecentNotes: %v", err)
+	}
+	for _, n := range notes {
+		if n.Source == EvalNoteSource {
+			t.Fatalf("wrote a note despite an LLM failure: %q", n.Text)
+		}
+	}
+}
+
+// TestParseObservationsTolerantAndBounded — Thinking models wrap JSON in prose,
+// and no reply may exceed MaxObservations even if the grammar is bypassed.
+func TestParseObservationsTolerantAndBounded(t *testing.T) {
+	obs, err := parseObservations(`<think>hmm</think> вот: [{"observation":"a","confidence":0.9,"suggested_action":"note"}] всё`)
+	if err != nil {
+		t.Fatalf("parse: %v", err)
+	}
+	if len(obs) != 1 || obs[0].Text != "a" {
+		t.Fatalf("got %+v, want one observation 'a'", obs)
+	}
+
+	var b strings.Builder
+	b.WriteString("[")
+	for i := 0; i < MaxObservations+3; i++ {
+		if i > 0 {
+			b.WriteString(",")
+		}
+		b.WriteString(`{"observation":"x","confidence":0.5,"suggested_action":"note"}`)
+	}
+	b.WriteString("]")
+	obs, err = parseObservations(b.String())
+	if err != nil {
+		t.Fatalf("parse: %v", err)
+	}
+	if len(obs) != MaxObservations {
+		t.Fatalf("parsed %d observations, want the %d cap", len(obs), MaxObservations)
+	}
+}