ZERO EXTERNAL DATABASES · v0.9.29

agentmemory

Give your coding agent a memory that survives the session. 95.2% retrieval recall on LongMemEval-S with about 92% fewer input tokens than full-context replay, all on your own machine.

See it move Star27k

AS FEATURED IN

AlphaSignal180K technical subscribers

Linux Foundation backed

Position #19 · NEW 2026

Live upvote count

95.2%

RETRIEVAL R@5 · LONGMEMEVAL-S

92%

FEWER INPUT TOKENS PER SESSION

MCP TOOLS

130

REST ENDPOINTS

EXTERNAL DATABASES

1655

TESTS PASSING

The stack

Three layers.
No framework tax.

Built on the iii engine: every memory operation is a worker, a function, or a trigger. No external databases, queues, or vector stores. The entire runtime is one process.

Hooks

12 auto-capture hooks piped into every coding agent. Every tool call, every prompt, every stop becomes a compressed observation, stamped with its origin and agent.

Recall

Hybrid retrieval on the primary recall path: BM25, vector, and knowledge graph scores ranked together, reranked on device. Superseded versions stay out of results.

Consolidate

With an LLM provider key, consolidation runs on session stop: raw observations compress into semantic memories, duplicates merge, stale rows decay, audit rows record it all.

What's inside

Twelve things you did not want to build.

agentmemory is not a library or a vector store. It is a complete memory runtime: capture, recall, consolidate, observe, federate.

12AUTO-HOOKS
Capture everything
Every session start, prompt, tool call, and stop fires into the memory pipeline with no glue code. Install the plugin and capture begins.
54MCP TOOLS
Native MCP surface
memory_save, memory_recall, memory_smart_search, memory_sessions, governance, audit, export: the full surface behind a single MCP server. Saving a near-duplicate returns a similarTo hint instead of a second copy.
130REST ENDPOINTS
HTTP first
Every MCP tool has a REST twin under /agentmemory/*. Curl it, fetch it from the browser, or proxy it from your own agent.
BM25+ VECTOR + GRAPH
Hybrid recall
The primary recall path ranks lexical, semantic, and graph scores together, reranked on device. Superseded versions stay out of results while their history stays queryable.
5ORIGIN CHANNELS
Provenance built in
Every record carries write-time provenance: user, agent, tool, import, or shared. Pass agentId through save and recall to scope memory per agent.
AUTOCONSOLIDATION
Raw to semantic
Activates with an LLM provider key. Consolidation runs on session stop: observations compress into semantic memories, duplicates merge, stale rows decay with retention scoring, and audit rows record the sweep.
∞REPLAY
JSONL session import
Point agentmemory at a Claude Code JSONL transcript and it rehydrates the session, indexes it for search, and derives crystals and lessons from what it finds.
GRAPHEXTRACTION
Knowledge graph
Entities and relations extract from observations when an LLM provider key is set and graph extraction is enabled. Query with /agentmemory/graph. Visualize in the viewer. Temporal edges supported.
IDXLESSON RECALL
Lessons that resurface
Save a lesson once, recall it by relevance later. Lessons live in a dedicated BM25 index with confidence and recency reranking, with save, recall, and delete over MCP and REST.
MESHFEDERATION
Peer-to-peer sync
Register another agentmemory node and push or pull memories over authenticated HTTPS. A bearer token is required; no silent syncs.
MDOBSIDIAN EXPORT
Your notes, hydrated
Mirror memories to a sandboxed vault directory as frontmatter-tagged markdown, ready for Obsidian's graph view.
0EXTERNAL DBs
One process
Runs as a single Node process with zero external services. State lives on disk as JSON. agentmemory stop flushes indexes before exit, in Docker mode too.

Command center

Two UIs.
One memory runtime.

agentmemory ships a real-time viewer for your memories and an engine-level console for every function, trigger, and OTel span. Both are first-class, installed inline by the CLI on first run.

Ship-with viewer · port 3113

The agentmemory server auto-starts a real-time viewer on port 3113. No install, no config. Tabs refresh live as hooks fire, and any past session replays in place.

LIVE OBSERVATION STREAM · EVERY HOOK AS IT FIRES
SESSION EXPLORER · REPLAY ANY PAST SESSION
MEMORY BROWSER · FILTER BY PROJECT / TYPE / CONFIDENCE
KNOWLEDGE GRAPH VISUALIZATION · FORCE-DIRECTED
HEALTH DASHBOARD · HEAP / RSS / EVENT LOOP LAG

$ open http://localhost:3113

Ship-with viewer · port 3113

Live

Memory that types back.

agentmemory@localhost:3111

IDLE

VS.

Vs. the field.

Only the agentmemory number is ours, measured on LongMemEval-S and reproducible from the repo. Competitor figures are their own published claims on their own benchmarks — different datasets, shown for ballpark. Ship what you want; we just picked the one with receipts.

AGENTMEMORYMEM0LETTAZEP / GRAPHITI

RETRIEVAL95.2% (LongMemEval-S)68.5% (LoCoMo)83.2% (LoCoMo)63.8% (LongMemEval)

EXTERNAL DEPS0Qdrant / pgvectorPostgres + vectorNeo4j

REST ENDPOINTS130———

MCP TOOLS54———

AUTO-CAPTURE HOOKS12Manual add()Agent self-edits—

NATIVE AGENT PLUGINS6———

OPEN SOURCEYes (Apache-2.0)YesYesYes

BUILDERS USING AGENTMEMORY

In the wild.

Verbatim from the Product Hunt launch thread. Each card links back to the source comment.

HOW THEY USE IT

Backfilled a month of Cursor transcripts

I backfilled agent memory on my past month's Cursor agent transcripts. It was surprisingly accurate. Picked up on things that I moved away from.

Peter NeyraProduct Hunt ↗

Two weeks of production use

Been using it for 2 weeks, and I definitely see improvements.

Pranav PrakashProduct Hunt ↗

WHAT THEY SAY

Tackles one of the biggest pain points with coding agents: losing useful project context across sessions without bloating the context window.

Alper TayfurProduct Hunt ↗

The focus on making memory actually useful for agents instead of just storing context endlessly.

Mia TaylorProduct Hunt ↗

Memory often becomes just more noise over time. Agentmemory feels more intentional compared to a lot of tools in this space.

Thomas HallProduct Hunt ↗

Tried it briefly — feels clean and easy to get started with.

Zoe AlexandraProduct Hunt ↗

Works with

Seven native plugins.
Rest MCP-native.

Native plugins for Claude Code, Copilot CLI, Codex CLI, OpenClaw, Hermes, pi, and OpenHuman. OpenCode gets a plugin that attributes each session to its own project. Every other MCP client gets it for free. `agentmemory connect <agent>` auto-wires them all.

FIRST-CLASS PLUGINClaude CodeFROM Anthropic

12 hooks + MCP + skills

NATIVE PLUGINCopilot CLIFROM GitHub

MCP + plugin hooks and skills

NATIVE PLUGINCodex CLIFROM OpenAI

6 hooks + MCP · native plugin

GATEWAY PLUGINOpenClawFROM openclaw

onSessionStart · onPreLlmCall · onPostToolUse · onSessionEnd

FIRST-PARTY INTEGRATIONHermesFROM Nous Research

Python plugin · yaml config

NATIVE PLUGINpiFROM pi

Native plugin + MCP

NATIVE BACKENDOpenHumanFROM tinyhumansai

Native Memory trait backend (Rust)

NATIVE PLUGINOpenCodeFROM SST

Capture plugin with per-session project attribution

Ship it

One install.
Any agent.

Runs on your machine. Data stays local. Capture and recall need no LLM key; add one for Anthropic, OpenAI, Gemini, MiniMax, or OpenRouter to activate consolidation, graph extraction, and LLM compression.

1. INSTALL ONCE

2. START THE MEMORY SERVER

3. RUN THE DEMO

ZERO-INSTALL PATH: NPX

4. WIRE UP ANY AGENT

One MCP JSON fits almost everything. Pick your agent on the left, or paste the universal config on the right.

AGENTS

CursorOPEN VS CodeOPEN

Cursor / VS Code are one-click via deeplink. Others copy the right snippet directly to your clipboard.

UNIVERSAL MCP JSONWORKS FOR CLAUDE DESKTOP · CURSOR · CLINE · ROO · WINDSURF · GEMINI · WARP · DROID · KIRO · ANTIGRAVITY · QWEN · MERGE INTO EXISTING mcpServers

{
  "mcpServers": {
    "agentmemory": {
      "command": "npx",
      "args": ["-y", "@agentmemory/mcp"],
      "env": {
        "AGENTMEMORY_URL": "http://localhost:3111"
      }
    }
  }
}

Read the quickstart npm package Integrations

agentmemory

Hooks

Recall

Consolidate

Capture everything

Native MCP surface

HTTP first

Hybrid recall

Provenance built in

Raw to semantic

JSONL session import

Knowledge graph

Lessons that resurface

Peer-to-peer sync

Your notes, hydrated

One process

Ship-with viewer · port 3113

In the wild.