Why this is in the vault
A clean explainer of agent memory (four memory types, the retain/recall/reuse/refine loop, and why forgetting is the unsolved part) that names the exact stale-fact failure our own file-based memory and vault are exposed to.
The core argument
LLMs are stateless; "memory" is an external system that writes notes and hands them back. The piece frames memory engineering as a loop (retain, recall, reuse, refine, a framing credited to Richmond Alake of Oracle) and says most systems handle the first three but break on refine, i.e. keeping memories true.
- Four types borrowed from the CoALA framework: working (context window), episodic (what happened), semantic (facts), procedural (rules, e.g. CLAUDE.md). Each goes stale at a different rate, so uniform treatment misbehaves.
- History: copy-paste, RAG, MemGPT/Letta tiered memory (Oct 2023), ChatGPT memory list (Feb 2024), then "dreaming" background rewrite passes (Anthropic for developer agents in May 2026, OpenAI for ChatGPT in June).
- Forgetting is not deleting: expiry dates on facts, fading by recency/importance, and a review pass that checks memories against the real source of truth. Evidence cited: STALE benchmark (best model 55.2%) and TEPA, where append-only memory (0.21) scored below no memory (0.31).
- Build vs buy: for a single user, a folder of text files plus an index, in git, goes far; buy a memory layer (Mem0, Zep, Letta) when many users need it; regulated work needs provenance. Portability across models is a stated requirement. The author's own anecdote: an agent reported a contract as pending days after it was signed.
Bias note: no sponsor block in this issue. The author promotes his own open source agent-memory template, and benchmark figures are cited second-hand via links (not verified here).
Mapping against Ray Data Co
Strong: this is a description of the MEMORY.md plus vault setup Ray already runs (index file read at startup, one file per topic, procedural rules in AGENTS.md), and it names the gap we manage by hand: refine. The AGENTS.md precedence rule "memory (newer wins; log the conflict in Errors & Corrections)" and the "verify blockers at source, not notes" feedback are manual versions of the article's "check memories against the real source" review pass. The concrete takeaway is a candidate scheduled memory-review pass that checks status-type memories (project_* entries, open blockers) against their source and proposes expiry or replacement with evidence, since dated facts such as "since 2026-09-24" go stale fastest. It also validates the file-plus-git choice for a single-founder setup and flags the "regulated/audited needs provenance" line as the point where it would not hold for client work.
Related
- [[2026-10-25-alphasignal-memorytax-agent-memory-study]]
- [[2026-07-28-alphasignal-optmem-agent-memory]]
- [[2026-04-10-akshay-pachaar-agent-harness-anatomy]]
- [[2026-02-23-every-chatgpt-memory-context-rot]]
- [[feedback_verify_blockers_against_source_not_notes]]