ANAMNESIS // MEMORY ATTENTION PIPELINE

How an agent remembers, and recalls.

The full sequence — write-side gating, then read-side retrieval across three nested layers.

ALWAYS-ON

OPT-IN

G1–G6 = the six attention mechanisms

DIAGRAM 1 — WRITE PATH · remember()

Embed content Always-On

384-dim vector via sentence-transformer

Salience Gate G6 Opt-In

surprise = 1 − exp(−nearest-neighbor distance) (embedding novelty)

salience = 0.5·(importance/10) + 0.2·emotion + 0.3·surprise

decision: importance ≥ 8 → CONSOLIDATE (hard floor) · salience ≥ 0.35 → CONSOLIDATE · else → EPHEMERAL (valid_until +14d)

Store Always-On

INSERT into memories + vector index + FTS index

The write side decides what's worth keeping — surprise + emotion + importance — never hard-dropping load-bearing facts.

DIAGRAM 2 — READ PATH · recall()

G5 FEDERATION — for own DB, then each peer DB (read-only) Opt-In

G1 MULTI-HOP — repeat up to N hops Opt-In

CORE recall() — per-query ranking

Embed query Always-On

384-dim vector representation of the query

PARALLEL SEARCH

Vector search Always-On

sqlite-vec · top 2k results · cosine similarity

FTS5 keyword search Always-On

BM25 ranking · top 2k results · token matching

RRF Fusion Always-On

score = Σ 1/(60 + rank) across both result lists

Validity gate Always-On

Drop expired records — valid_until passed

RANK — two routes, one wins

ROUTE A — short-circuits B

Cross-encoder rerank G4

Opt-In

Score query × content relevance, sort descending. Wins — short-circuits Route B.

ROUTE B — default when A absent

× importance weight Always-On

Scale by importance factor 0.6–1.5

× recency decay G3 Opt-In

0.5^(age / half-life)

+ emotion rerank Opt-In

cosine(query_emotion, memory_emotion) · 0.3

Sort + trim to limit Always-On

Final sort descending on composite score · slice to requested limit

Associative expansion G2 Opt-In

Pull neighbors sharing association labels · score by overlap × importance · append with provenance — spreading activation

CORE RECALL — per-query fusion · vector + lexical · filtered · ranked

assess coverage (Haiku or heuristic) → refine query → loop · stop on: assessor-stop · budget · convergence

merge peer results · dedup · tag source_scope: own → peer:<agent>

MENTAL MODEL

Federation sets the breadth · multi-hop sets the depth · the core fuses semantic + lexical evidence · the rerankers set priority · expansion follows the association graph outward. Write-side, the salience gate decides what was worth admitting at all.