agent-memory-curator
The Agent Memory Curator consolidates session logs into structured, dated memory entries by reading episodic notes, distilling them into one-line facts, and reconciling contradictions by updating existing entries rather than duplicating them. Use this skill at the end of a working session or after resolving a memorable incident to promote vetted observations from logs into MEMORY.md and semantic memory files while filtering for PII and applying distillation rules.
git clone --depth 1 https://github.com/oxbshw/LLM-Agents-Ecosystem-Handbook /tmp/agent-memory-curator && cp -r /tmp/agent-memory-curator/skills/examples/agent-memory-curator ~/.claude/skills/agent-memory-curatorSKILL.md
# Agent Memory Curator ## When to use - End of a working session - After resolving an incident worth remembering - When `MEMORY.md` has grown stale or contradictory ## When NOT to use - During a session (don't write memory mid-flight from low-confidence signals) - For raw transcripts (those stay in `logs/`) ## Inputs | Name | Type | Required | Notes | |---|---|---|---| | `since` | date | no | default: last curation time | | `categories` | list | no | restrict to e.g. `[user, project, decision]` | ## Workflow 1. **Read** new episodic logs in `logs/episodic/` since `since` 2. **Distill** each notable event into a one-line dated entry 3. **Categorize** into: user / project / decision / feedback / open question 4. **Reconcile**: if a new entry contradicts an existing one, *update* the existing one rather than appending; move the old fact to "Outdated" 5. **Apply rules** from `references/memory-distillation-rules.md` 6. **Write** to `MEMORY.md` (index) and `memory/semantic/<topic>.md` (details) 7. **Verify** no PII / secrets leaked into memory 8. **Report** summary of additions / updates / deletions ## References - [`references/memory-distillation-rules.md`](references/memory-distillation-rules.md) ## Success criteria - Every new entry is dated - No duplicate entries (search before adding) - No PII / secret values - Outdated entries moved, not deleted in-place ## Failure modes - Conflicting evidence → mark as "Open question" rather than guessing - Source untrusted (e.g., from MCP/web fetch) → require human confirm before promotion
Use when capturing an architecture decision so it survives turnover — produces an ADR-NNNN.md from context, options considered, and the chosen path.
Use when reviewing a proposed REST or GraphQL API change before merge — checks contract clarity, backwards compatibility, errors, pagination, auth, and naming.
Use when first encountering a new dataset — produces a structured profile (schema, missingness, distributions, outliers, gotchas) before any analysis.
Use after an incident is resolved — drafts a blameless postmortem from timeline notes, alerts, and chat threads.
Use when opening a PR — produces a clean PR description (what / why / how to verify / risks) from a branch diff against base.
Use when planning the next sprint — turns ticket intake + team capacity into a planned sprint with explicit non-goals.
Use before connecting a new MCP server to your agent — produces a structured security review covering source, permissions, tools, network, and approvals.