Agent Memory

Give AI agents persistent memory that survives across conversations. Store structured memories (decisions, learnings, patterns, preferences, facts) with confidence scores, search by natural language, and retrieve contextually relevant memories. ## What It Does - **Memory Search** — Natural language search ranked by confidence and FTS relevance. Filter by memory kind. - **Contextual Recall** — Two-stage retrieval with LLM reranking for topic-specific memory recall. - **Memory Creation** — Store structured memories with kind, importance, tags, and confidence scores. - **Memory Correction** — Fix a stored memory in place, including the one-line summary recall shows beside its title, or delete one that turned out wrong. Only the fields you send change. - **Statistics** — Overview of memory collection: totals by kind, importance distribution, average confidence. - **Lessons Near the Code** — Before an agent changes a file, it reads the lessons that apply to that code. `entryContext` finds the ones on the file, its functions and its folders. `entriesNear` adds the ones on code the file imports, calls or sits next to. `anchorsLookup` counts the lessons on each folder, file or function. `entriesSimilar` checks if a lesson exists before the agent writes it twice. When a file is renamed, its lessons move with it. - **Rules Checked by a Test** — A lesson can name the test that checks it, on a line that starts with `Checked by:`. It can also name the incident it came from, with `Derived from:`. When every test a lesson names is deleted, the lesson is marked unguarded, and one search lists all such rules. A renamed test keeps guarding. `entriesNear` puts rules with no nearby test first. `entryContext` on a test file shows the rules it protects. ## Best For - Agents that need to remember user preferences across sessions - Storing decisions and learnings from past interactions - Building institutional knowledge that persists beyond conversations - Context injection for informed agent responses