Agent Memory MCP — Kaustubh Mani Tripathi

Not runnable on this website. The MCP server and SQLite daemon need a local process (or a future isolated host). This page is a project overview; install and run on your machine.

What it does

Agent Memory MCP watches your workspace, stores failures, decisions, and attempts in SQLite, and exposes search and context tools to any MCP-compatible client (Cursor, Claude Desktop, Zed, etc.). Optional LLM extraction adds facts and preferences without shipping code to the cloud by default.

Architecture

Watcher / MCP / HTTP → Storage → MemoryEngine (SQLite)
                              ↓
                    Optional LLM extractor → facts

MCP tools

Tool	Purpose
`remember`	Store memory for a project path
`search_memory`	Keyword + semantic search (FTS5 + local vectors)
`get_repo_context`	Bundle failures, decisions, attempts for a repo
`mark_failure_resolved`	Clear a recurring failure signature
`forget`	Soft-delete by id or signature

Prompt: inject_memory_context — preload repo memory into a session by path.

Run locally

cd agent_memory_mcp
pip install -e ".[dev]"
# Optional LLM extraction:
pip install -e ".[llm]"

# Daemon (API :9000, watcher on by default)
agent-memory serve --root ~/agent_companion_data

# MCP client config (example)
# command: agent-memory  args: ["mcp", "--root", "~/agent_companion_data"]

Configure via mcp.json in your editor; see the project README for AGENT_MEMORY_LLM_PROVIDER and NVIDIA/Hermes auth notes. Data lives under ~/agent_companion_data/memory.db — not on this static site.

Operator UI

A small dashboard (memory browser, logs, configuration) ships with the repo and talks to the same backend on port 9000. Build with npm run build in ui/, then agent-memory serve --ui.

How it works — components

Filesystem watcher

Passively observes the workspace for log errors and code churn — the theory is that the most useful memory is captured automatically, not when an agent remembers to write it.

SQLite + FTS5 + vectors

A single embedded DB holds structured rows; full-text (FTS5) plus local vectors give hybrid keyword+semantic recall without a cloud vector store.

MCP tool surface

remember, search_memory, get_repo_context expose memory to any MCP client — the protocol boundary keeps the store editor-agnostic.

Optional LLM extractor

When enabled, an LLM distills raw events into facts/preferences; default-off keeps the system local and cheap, extraction is an opt-in enrichment layer.

Elsewhere

Source: local project agent_memory_mcp under localcode/. Profile: github.com/xdutsuay for related experiments.

← All projects