Seven Kinds of Agent Memory, Checked
A seven-part taxonomy of agent memory has been circulating, and it deserves the compliment this series reserves for its rarest verdict: it is substantially correct. The seven categories map onto how the agent-memory literature and shipping products actually divide the problem. What the poster undersells is the cost side — every kind of memory you add is a new way for your agent to be confidently wrong about stale facts.
01 — The seven, restated and checked
- Working memory (in-context) — everything currently in the context window: system prompt, messages, tool outputs. Correct, and correctly labeled short-term. This is the only memory that needs no engineering.
- Semantic memory — persistent facts and preferences, decoupled from when they were learned. Matches product "memory" features and the research term.
- Episodic memory — a log of specific past events and runs, so the agent can learn from what worked and failed. Correct; in practice this is transcripts, run journals, and lesson files (the compounding trick in A-05).
- Procedural memory — how-to knowledge: skills, tool patterns, behavioral rules. Correct, and in current systems it has a concrete address: skills and instruction files (C-11).
- External / retrieval memory — knowledge fetched at inference time by similarity search. Correct: this is RAG (A-07), classed properly as a memory type.
- Parametric memory — what's baked into the weights at training. Correct, and the honest poster admits you don't control this one; fine-tuning is its only lever.
- Prospective memory — remembering future intentions: scheduled goals, planned-but-not-executed tasks. The most forward-leaning label, but real: scheduled tasks and goal-conditioned runs are exactly this, shipping today.
02 — The build order, keptStart small; every layer is a liability
The poster's practical takeaway is genuinely good engineering advice and we keep it verbatim in spirit: start with working memory; add semantic memory when users expect to be remembered; layer episodic, procedural and prospective only when the agent must plan ahead, learn from failure, and adapt. The addition it needs: each layer also adds a failure mode — stale facts served confidently, old lessons applied to new situations, retrieval surfacing the wrong document with authority. Memory without expiry dates and provenance is how agents become confidently wrong. Budget maintenance, not just storage.
03 — The ledgerChecked 2026-08-06
| Claim | Verdict | The record |
|---|---|---|
| The seven-category taxonomy itself | VERIFIED | Aligns with the cognitive-architecture framing used in agent-memory research and products. |
| Retrieve → reason → act → learn as the cycle memory serves | VERIFIED | Standard agent loop, correctly positioned. |
| Working memory = the context window | VERIFIED | Exactly right, including its short-term nature. |
| RAG classed as a memory type | VERIFIED | Fair and useful framing. |
| Prospective memory as a distinct, practical class | NUANCE | Real (scheduled goals ship today); the research label is newer than the poster implies. |
| Incremental build order (start with working memory) | VERIFIED | Sound engineering; adopted above. |
| Memory as pure upside | NUANCE | The poster never says it — but never prices staleness, either. Every layer is also a liability. |