Agent memory and recall
Overview
A session ends and the conversation context dies with it. What survives is what was written down — and the difference between an agent that repeats its mistakes every session and one that compounds its knowledge is the discipline of the write-down. Persistent memory is not a product you install; it is a practice: what to record, in what shape, promoted how, superseded when, with what review triggers.
This skill covers that practice. Its neighbours hold the complementary mechanics: ac-prompt-engineer's tokens-capacity leaf budgets the window within one session; ac-retrieval-and-vector-systems builds the search layer that memory sits on; ac-handoff resumes one interrupted task. This skill owns the durable layer — what persists after the session is gone — and the recall discipline that makes the persisted layer usable.
Primary sources: https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents and https://en.wikipedia.org/wiki/Knowledge_management.
When to use
- When an agent's knowledge must survive session boundaries.
- When designing what gets written down at the end of work.
- When memory has grown and needs cleanup or supersession.
- When retrieved memories are trusted more or less than they should be.
When NOT to use
- For in-window context budgeting and compaction (use ac-prompt-engineer).
- For index, chunking, and embedding mechanics (use ac-retrieval-and-vector-systems).
- For the mid-task checkpoint that resumes interrupted work (use ac-handoff).
Subskill Index
| Leaf | Domain segment | Scope (one line) |
|---|---|---|
subskills/what-persists.md |
prompt-engineering/what-persists |
The session-end write-down: what outlives the session and what dies with it. |
subskills/note-anatomy.md |
prompt-engineering/note-anatomy |
One fact per note, dated, titled for retrieval, tagged for coverage. |
subskills/promote-and-supersede.md |
prompt-engineering/promote-and-supersede |
Working note to permanent memory, dedupe, and contradiction handling. |
subskills/recall-discipline.md |
prompt-engineering/recall-discipline |
Reading memory before acting, matching query to note, and trust calibration. |
subskills/staleness-and-review.md |
prompt-engineering/staleness-and-review |
Lifecycle states for facts, review cadences, and pruning. |
subskills/knowledge-manage.md |
prompt-engineering/knowledge-manage |
Memory cleanup, archive versus delete, and governance of the whole corpus. |
subskills/migration-and-portability.md |
prompt-engineering/migration-and-portability |
Moving memory when infrastructure dies; format and export discipline. |
subskills/memory-record.md |
prompt-engineering/memory-record |
The inventory of what is remembered, reviewed on a cadence. |
Status and honest limits
Memory is a liability when stale: an agent that recalls a deprecated fact with confidence is worse than one that recalls nothing. Every technique here assumes the write-down is cheap and the recall is searchable; where retrieval infrastructure is absent, the discipline still applies to plain files. Nothing here predicts what a model will hold in its context window — that is in-window mechanics — and no storage product is recommended.