aegisdb-mcp

# AegisDB ↔ Claude Code Memory Integration Make [AegisDB](https://github.com/d4n-larsson/aegisdb) the **persistent long-term memory** of Claude Code. The agent gets memory **tools** (save/search/get/update/relate) via an MCP server, plus automatic **recall** and **capture** via hooks — so knowledge learned in one session (decisions, conventions, fixes, preferences) is available in later ones without the user re-explaining it. Each project keeps its own isolated memory. ## How it works ``` Claude Code ──(MCP stdio)──▶ aegis_mcp.server ──┐ Claude Code ──(hooks)──────▶ recall/capture ────┼──▶ AegisDB (NDJSON/TCP) embeddings ─────────┘ ``` - **MCP tools** (`mcp__memory__memory_save`, `_search`, `_get`, `_update`, `_relate`) — explicit, model-driven memory. - **`UserPromptSubmit` hook** — automatic recall: injects relevant memories into context before each turn, best-effort under a time budget. - **`SessionEnd` hook** — automatic capture: persists salient session outcomes. - **Embeddings** — pluggable provider (Voyage / local / none) turns text into vectors for semantic recall; the integration never asks the agent for vectors. All logic lives in dependency-free modules under `aegis_mcp/`; only the MCP server entry point needs the `mcp` SDK. Memory is always best-effort: if AegisDB is down, the agent stays fully usable. ## Why it saves tokens Long context is the real cost driver, and this integration keeps durable knowledge **out** of the window — feeding back only what's relevant per prompt — so you spend tokens on the work, not on re-establishing context. - **Recall instead of re-explaining.** Stack, conventions, decisions, and gotchas learned earlier are injected automatically, so you stop re-pasting them every session and the model stops re-deriving them. - **A relevant slice, not a dump.** Recall ranks by similarity × importance × confidence and injects only the top matches — capped by `AEGIS_RECALL_TOP_K`, filtered…