Evidence

AideMemo is designed around a local, retrieval-first memory loop. The default memory path covers capture, typed writes, BM25-first search, and MCP or agent SDK reads, and does not require an external LLM call. Remote extraction, embedding, reranking, and reader models remain opt-in.

This page summarizes the results that currently shape product defaults. See the full Measurement Ledger for commands, fixtures, caveats, and historical runs.

Validated outcomes

Surface and result	What it supports
LongMemEval-S retrieval, opt-in BGE plus two-stage rerank, 500 questions R@10 0.992, MRR 0.958	The semantic path can recover paraphrase-heavy memory when lexical retrieval is not enough.
LongMemEval-S end to end, same retrieval plus MiniMax reader 74.0%	Reader-backed evaluation is competitive enough to study, but this is not a default-path or SOTA claim.
BrainBench, BM25 via daemon P@5 17.4%, R@5 64.1%; same score and 5.7x faster than fresh CLI	Keep surface-form-heavy search lexical and keep the store warm.
Shared HTTP MCP, 2 clients x 10 writes 20/20 persisted; p50 18.4ms, p95 41.8ms	A single local daemon is the preferred concurrent writer path.
Zero-token workflow demo decision, lesson, and error surfaced in 128ms	The core workflow can be demonstrated without an agent or model call.
Cross-agent handoff Scenario P 12/12 quality gates; critical evidence 4/4, route 4/4, neighbouring-source leakage 0; structured SDK packet and done_when preserved; handoff context -82.8% vs raw thread and -36.1% vs session canvas	An orchestrator can route one tracked workflow across agent/profile boundaries with bounded, fact-linked context, an observable completion condition, and a one-command receiver resume. This proves the deterministic artifact contract, not downstream model task success.
Multi-account handoff Scenario Q 11/11 gates across codex-one, codex-two, and claude-main; codex-one -> codex-two -> codex-one round trip; actor/source leakage 0; dispatch creates one pointer entity and zero copied facts; broker/payload keys 0	Account installations can bidirectionally pull and acknowledge the same tracked session without sharing vendor-local chat ids. This is a pointer ledger, not evidence of authentication, queue delivery, exclusive ownership, or downstream task success.
Hermes Kanban boundary Scenario R 13/13 gates with a real temporary Hermes Kanban DB; board/tenant scope and profile actor derived from dispatcher metadata; internal coding -> reviewer creates zero AideMemo assignments; external codex-two dispatch adds one pointer and zero facts; same-session evidence returns before Hermes explicitly marks the card done	Kanban remains the canonical task lifecycle while AideMemo carries durable evidence across an external installation boundary. The derived source is trusted-process routing, not gateway authentication. This does not prove an external CLI worker spawner or downstream model task success.
External worker lane Scenario S 16/16 fake Codex/Claude gates; automatic workers persist distinct exclusive claim tokens and a failed assignment accepts a new worker claim; success receives the packet and resume environment, returns same-session evidence, then completes; failure records a same-session error and remains accepted; sender outbox/status link both facts	The packaged receiver can execute the handoff/return protocol with shell-free argv while leaving Hermes Kanban untouched. This does not prove live-model task success, authentication, automatic retry, exactly-once execution, or Hermes spawn_fn integration.
Agent SDK package smoke wheel install plus Memory, client, worker-lane exports, and installed aidememo-worker-lane --help checks passed in 3.28s	The code-first integration and external receiver are independently packageable, not tied to one agent runtime.

These measurements have different datasets and execution envelopes. Compare rows within their stated benchmark, not as one aggregate score.

Model placement

Failure point and current placement	Evidence boundary
Normal code and docs search BM25-first search.auto_hybrid=true, multilingual model2vec semantic fallback, daemon prewarm	BrainBench stayed quality-equivalent on the lexical daemon path while avoiding fresh-process overhead.
English paraphrase-heavy memory Opt-in bge-small-en-v1.5	LongMemEval-S R@5 improved from 96.2% to 98.0%; the roughly 10x warm query cost is not justified for every workload.
Weak first-stage lexical recall Guarded MLX LFM embedding experiment	On 180 agent-trace documents and 540 queries, BM25 R@8 was 0.991, pure LFM dense was 0.887, and guarded auto reached 0.993 while promoting 2 weak cases. LFM is not the global default embedding replacement.
Good candidates with poor ordering Warmed LFM ColBERT experiment	A tiny fixture improved hit@1 from 0.57 to 0.86; candidate recall, document-token cost, and a larger corpus gate must be proven before product placement.
Missing or ambiguous fact type LFM 1.2B LoRA shadow hint	At confidence >= 0.98, the expanded high-signal trace gate accepted 39/155 hints at precision 0.923 with 0 baseline-correct harms. It remains review data, not an automatic write decision.
Privacy-sensitive writes Opt-in local MLX privacy sidecar plus deterministic secret prefixes	The measured MLX sidecar reduced warm write overhead relative to the CPU model, but memory and latency costs still make explicit enablement the honest default.

Claim boundaries

• AideMemo is a memory and retrieval system, not a hosted agent runtime.
• The default memory loop does not call an external LLM. Opt-in extractors, TEI endpoints, rerankers, and benchmark readers can.
• Small local models are promoted only where a scenario gate shows a useful quality and latency trade-off. A neutral result keeps the cheaper path.
• LongMemEval results calibrate retrieval and reader behavior; AideMemo does not lead with a state-of-the-art claim.
• Registry publication status is tracked separately in the Release Checklist.

For the system boundary behind these paths, continue with Architecture.