# Agent Memory

> Use when: persist and recall context across coding sessions so the agent stops re-learning the same stack, bugs, and decisions.

- Skill: `kimtth/agent-memory` (Agent Skill)
- Install (CLI): `npx skillmds@latest add kimtth/agent-memory`
- Raw SKILL.md: https://api.skillmd.com/api/skills/kimtth/agent-memory/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- License: Apache-2.0
- Author: kimtth (https://skillmd.com/u/kimtth)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/kimtth/agent-memory

---


Goal: give the agent searchable long-term memory — capture what happened, compress it to facts, and inject the right context into the next session.

Use for:
- avoiding re-explaining architecture, preferences, and prior decisions each session
- recalling why a choice was made (e.g. jose over jsonwebtoken for Edge compatibility)
- handing off context between agents or teammates

Pipeline:
1. Capture observations from tool use; strip secrets and API keys before storing.
2. Compress raw observations into structured facts, concepts, and a short narrative.
3. Consolidate across four tiers:
   working (raw) -> episodic (session summaries) -> semantic (facts) -> procedural (workflows).
4. Index for hybrid search: BM25 keywords + vector similarity + knowledge-graph traversal,
   fused with reciprocal rank fusion.
5. On session start, load the project profile and inject top results under a token budget.

Lifecycle:
- Memories decay over time; frequently accessed ones strengthen, stale ones auto-evict.
- Detect and resolve contradictions; trace any memory back to its source observation.

Operations:
- remember (save) | recall / smart-search | session-history | handoff | forget

Rules:
- Inject a bounded budget (~2k tokens), not the whole store — keep context lean.
- Keep capture passive and privacy-filtered; never persist credentials.
- Memory augments reading current files; it does not replace verifying them.