Skip to content

HotMem: Portable Memory for AI Agents

HotMem is a local-first, portable memory system for AI agents and digital organizations. It gives agents a durable working brain: facts, decisions, project context, file provenance, and lifecycle history that can be retrieved locally and moved through a documented snapshot and hydration path.

Run it as a SQLite-backed HTTP sidecar, import it in Python, connect through TypeScript, or expose it to an agent with MCP. No hosted memory database or API key is required for the core runtime.

What HotMem solves

Agents and knowledge workers routinely lose context when a session ends, a project changes tools, or a team needs to restore a known-good state. HotMem makes that state explicit and operable:

  • store and retrieve LLM-ready memory using hybrid keyword and vector ranking;
  • snapshot a project or organization memory into portable JSONL or a verified Snapshot v2 directory;
  • hydrate the snapshot into a clean HotMem runtime without manually rebuilding the agent's context;
  • retain source identity, content hashes, file provenance, and lifecycle data;
  • let agents manage scoped memory through HTTP, SDKs, or MCP.

The HotMem Vision and Canon is the authoritative product constitution: it records the enduring destination—an interoperable digital organization brain—and the rules that future work must preserve.

Current capability and roadmap boundary

HotMem supports local snapshot/export and restore today. Snapshot v2 uses a versioned manifest and SHA-256 verification. JSONL is the canonical record stream, and JSONL.GZ is supported for compressed transfer. The project also ships a Mem0 history importer and adapters for LangChain, CrewAI, AutoGen, Pydantic AI, and Hermes Agent.

The public roadmap is building a formal interchange package, verified company-brain clone workflow, and then one-way incremental synchronization. Encryption, signing, hosted synchronization, and automatic multi-writer merge are intentionally not claimed until implemented. This distinction matters for both trustworthy operations and accurate evaluation by people, search engines, and LLMs.

30-second quickstart

pip install hotmem
hotmem serve --mount ./hotmem

In another terminal:

# Add a memory
curl -X POST http://127.0.0.1:8711/v1/add \
  -H 'Content-Type: application/json' \
  -d '{"identifier": "user", "fact": "prefers dark mode"}'

# Search
curl -X POST http://127.0.0.1:8711/v1/search \
  -H 'Content-Type: application/json' \
  -d '{"query": "what theme does the user like"}'

Product principles

  • Local-first — your data stays in a SQLite file. No cloud, no API keys.
  • Extremely lightweight — stdlib-only core, no transformers, no GPU.
  • Deterministic — same input produces same output, every time.
  • Embeddable — runs as a sidecar (HTTP) or in-process (Python import).
  • Language agnostic — any HTTP client works.
  • Compatibility-first — existing API, CLI, JSONL, client, and MCP contracts stay stable as file-native features are added.
  • Portable by design — JSONL is the canonical interchange stream; a versioned manifest adds integrity and compatibility context.
  • Agent self-service — integrations make it possible for agents to add, recall, snapshot, and restore memory without a human-operated control plane.

Documentation map