Memory hygiene: a relevance floor, no duplicates, and expiry that actually deletes

Three things that decide whether memory improves with use or rots with it.

A RELEVANCE FLOOR. Recall returned its top five whatever they scored, so a run
about shift cover was handed five memories about certifications simply because
nothing better existed — and the block tells the model these are things the
workspace remembered, so it reads them as pertinent. Embeddings are
unit-normalised, so knowledge_dot is cosine, and 0.30 is where text is usually
about something else. A judgement rather than a measurement, and the honest way
to tune it is to watch what gets carried on real questions.

NO DUPLICATES. The same standing preference comes up in conversation after
conversation, and each run that hears it has no idea the last one wrote it
down. Five recall slots spent on one fact restated five ways is the normal
failure, not a rare one. A write with the same normalised text, in the same org
and about the same subject, pushes the existing memory's expiry out instead of
adding a row — matched on the same sentence rather than a similar one, because
collapsing two genuinely different facts is the worse error.

EXPIRY THAT DELETES. expires_at was set and filtered on read, and nothing ever
removed anything: the row was invisible and still retained. "We keep it ninety
days" has to be true of the table, not only of the query. Prune is batched, and
a redaction is kept for a thirty-day grace period so an erasure stays provable
shortly afterwards.

It runs in the maintenance sweeper that already exists rather than a second
scheduler — same ticker, same cancellation, same failure isolation. That forced
one honest change: Maintenance() used to be nil without OAuth, on the reasoning
that there was nothing to sweep. There is now, and a retention promise enforced
only when an unrelated feature happens to be enabled is not a promise. The test
that asserted the old behaviour now asserts the new one and says why.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

go-api/internal/httpserver/maintenance.go

@@ -5,6 +5,7 @@ import (
 	"log/slog"
 	"time"
 
+	"github.com/krow/krow-backend/go-api/internal/memory"
 	"github.com/krow/krow-backend/go-api/internal/oauth"
 	"github.com/krow/krow-backend/go-api/internal/ratelimit"
 )
@@ -64,7 +65,14 @@ const maintenanceTimeout = 60 * time.Second
 type Maintenance struct {
 	store   *oauth.Store
 	limiter *ratelimit.Limiter
-	log     *slog.Logger
+
+	// memories is the long-term memory store, or nil where a deployment does
+	// not keep any. Swept here rather than on its own schedule: expiry is a
+	// retention promise, and a promise enforced by a second mechanism is one
+	// that can be switched off without anybody noticing.
+	memories *memory.Store
+
+	log *slog.Logger
 }
 
 // Maintenance exposes the sweeper, or nil when there is nothing to sweep.
@@ -72,14 +80,19 @@ type Maintenance struct {
 // Mirrors Server.Sessions(), which exists for exactly this reason: the process
 // owns the schedule, the server owns the things being swept.
 func (s *Server) Maintenance() *Maintenance {
-	if !s.cfg.OAuth.Enabled() {
+	/* Memory expiry has to run even where OAuth is off, so the nil check can
+	   no longer be about OAuth alone: a deployment that keeps memories and
+	   does not issue tokens would otherwise retain personal data forever
+	   because an unrelated feature was disabled. */
+	oauthOn := s.cfg.OAuth.Enabled()
+	if !oauthOn && s.memories == nil {
 		return nil
 	}
-	return &Maintenance{
-		store:   oauth.NewStore(s.db.Pool),
-		limiter: s.limiter,
-		log:     s.log,
+	m := &Maintenance{limiter: s.limiter, memories: s.memories, log: s.log}
+	if oauthOn {
+		m.store = oauth.NewStore(s.db.Pool)
 	}
+	return m
 }
 
 // MaintenanceResult is what one pass removed.
@@ -88,11 +101,12 @@ type MaintenanceResult struct {
 	AccessTokens  int64
 	RefreshTokens int64
 	RateLimits    int64
+	Memories      int64
 }
 
 // Total is the row count removed, for the log line.
 func (r MaintenanceResult) Total() int64 {
-	return r.Grants + r.AccessTokens + r.RefreshTokens + r.RateLimits
+	return r.Grants + r.AccessTokens + r.RefreshTokens + r.RateLimits + r.Memories
 }
 
 // Sweep runs one maintenance pass.
@@ -108,13 +122,26 @@ func (m *Maintenance) Sweep(ctx context.Context) (MaintenanceResult, error) {
 	// OAuth: codes, access tokens, and refresh tokens past their retention.
 	// The grace period and the reuse-detection retention are enforced inside
 	// Store.Cleanup — this schedules it, it does not reimplement it.
-	cleaned, err := m.store.Cleanup(ctx)
-	if err != nil {
-		firstErr = err
-	} else {
-		out.Grants = cleaned.Grants
-		out.AccessTokens = cleaned.AccessTokens
-		out.RefreshTokens = cleaned.RefreshTokens
+	if m.store != nil {
+		cleaned, err := m.store.Cleanup(ctx)
+		if err != nil {
+			firstErr = err
+		} else {
+			out.Grants = cleaned.Grants
+			out.AccessTokens = cleaned.AccessTokens
+			out.RefreshTokens = cleaned.RefreshTokens
+		}
+	}
+
+	/* Independent of the others for the same reason they are independent of
+	   each other: this is the sweep that keeps a retention promise, and a
+	   failure elsewhere must not be the reason personal data outlives it. */
+	if m.memories != nil {
+		pruned, err := m.memories.Prune(ctx, 0)
+		if err != nil && firstErr == nil {
+			firstErr = err
+		}
+		out.Memories = pruned
 	}
 
 	if m.limiter != nil {
@@ -170,11 +197,13 @@ func SweepMaintenance(ctx context.Context, m *Maintenance, log *slog.Logger) {
 		case err != nil:
 			log.Warn("maintenance sweep failed", "error", err,
 				"grants", result.Grants, "access_tokens", result.AccessTokens,
-				"refresh_tokens", result.RefreshTokens, "rate_limits", result.RateLimits)
+				"refresh_tokens", result.RefreshTokens, "rate_limits", result.RateLimits,
+				"memories", result.Memories)
 		case result.Total() > 0:
 			log.Info("maintenance sweep",
 				"grants", result.Grants, "access_tokens", result.AccessTokens,
-				"refresh_tokens", result.RefreshTokens, "rate_limits", result.RateLimits)
+				"refresh_tokens", result.RefreshTokens, "rate_limits", result.RateLimits,
+				"memories", result.Memories)
 		default:
 			log.Debug("maintenance sweep found nothing to delete")
 		}

go-api/internal/httpserver/maintenance_test.go

@@ -249,21 +249,34 @@ func TestMaintenanceSurvivesADatabaseFailure(t *testing.T) {
 	}
 }
 
-/* ── It is absent when the surface is ───────────────────────────────────── */
+/* ── It runs wherever there is something to retain ──────────────────────── */
 
-// A deployment without OAuth has nothing to sweep, and must not start a ticker
-// that runs for the life of the process doing nothing.
-func TestMaintenanceIsNilWhenTheSurfaceIsDisabled(t *testing.T) {
+// This used to assert the opposite: no OAuth meant nothing to sweep, so no
+// ticker. Long-term memory changed the premise. Memories carry an expiry that
+// is a retention promise about personal data, and a promise enforced only when
+// an unrelated feature happens to be switched on is not a promise. So the
+// sweeper now exists wherever the database does.
+func TestMaintenanceRunsForMemoryEvenWithoutOAuth(t *testing.T) {
 	a := newAPI(t) // the standard fixture: no OAuth configuration
 
-	if m := a.srv.Maintenance(); m != nil {
-		t.Error("an unconfigured deployment returned a Maintenance sweeper")
+	m := a.srv.Maintenance()
+	if m == nil {
+		t.Fatal("no sweeper, so expired memories would be retained forever")
 	}
 
-	// And the runner must return immediately rather than tick forever.
+	// It must still do a pass without OAuth configured rather than failing on
+	// the half that is absent.
+	if _, err := m.Sweep(context.Background()); err != nil {
+		t.Errorf("a sweep without OAuth failed: %v", err)
+	}
+}
+
+// And the runner still returns immediately when there is genuinely nothing,
+// rather than ticking for the life of the process.
+func TestSweepMaintenanceReturnsImmediatelyWithNothingToSweep(t *testing.T) {
 	done := make(chan struct{})
 	go func() {
-		httpserver.SweepMaintenance(context.Background(), a.srv.Maintenance(),
+		httpserver.SweepMaintenance(context.Background(), nil,
 			slog.New(slog.NewTextHandler(io.Discard, nil)))
 		close(done)
 	}()

go-api/internal/httpserver/server.go

@@ -37,6 +37,7 @@ import (
 	"github.com/krow/krow-backend/go-api/internal/db"
 	"github.com/krow/krow-backend/go-api/internal/definition"
 	"github.com/krow/krow-backend/go-api/internal/knowledge"
+	"github.com/krow/krow-backend/go-api/internal/memory"
 	"github.com/krow/krow-backend/go-api/internal/ratelimit"
 	"github.com/krow/krow-backend/go-api/internal/runtime"
 	"github.com/krow/krow-backend/go-api/internal/service"
@@ -60,6 +61,11 @@ type Server struct {
 	runs    *runtime.RunReader
 	version string
 
+	// memories is the long-term memory store, for the scheduled sweep that
+	// enforces its retention promise. The runtime builds its own; this one is
+	// here so maintenance can prune without reaching through the engine.
+	memories *memory.Store
+
 	// toolCatalogue is the tool set an agent author may choose from.
 	//
 	// Built whether or not a model credential exists: the catalogue describes
@@ -254,6 +260,11 @@ func New(cfg *config.Config, database *db.DB, log *slog.Logger, opts ...Option)
 		s.runs = runtime.NewRunReader(database.Pool)
 	}
 
+	// Memories are swept wherever the database is, agents or not: a deployment
+	// that stops serving agents still holds what earlier ones remembered, and
+	// retention is a promise about the table rather than about the feature.
+	s.memories = memory.New(database.Pool, runtime.NewEmbedder(*cfg))
+
 	// Built the same way the runtime builds its own, so the list an author is
 	// offered is the list their agent will actually have.
 	toolRegistry := runtime.DefaultTools(