neatmem
kanhaoning/neatmem
NeatMem long-term memory plugin for DeepSeek Harness — thin REST client against a neatmem server
Install
dsh plugin --profile web add github:kanhaoning/neatmemREADME
Lightweight local memory for agents — every dedup, update, and rerank decision inspectable and tunable.
3 dedup detectors × 4 update resolvers × 3 rerank modes · 60+ parameters · 6 prompts replaceable
Benchmark
LoCoMo accuracy (5-run mean, MiniMax-M3) · reproduction
| Question type | Accuracy |
|---|---|
| single-hop | 92.4% |
| temporal | 93.5% |
| multi-hop | 90.0% |
| open-domain | 69.8% |
| Overall | 90.8% |
Why NeatMem?
Agent memory is easy to start but hard to keep clean.
Common problems include:
- duplicate memories accumulating over time
- semantically related memories not being merged
- irrelevant memories being recalled because of weak vector matches
NeatMem keeps every memory decision inspectable and tunable:
- Every extraction, dedup, merge, and rerank decision runs through a prompt you can read and replace — 6 prompt slots, as plain text files.
- Every threshold and behavior switch is an explicit parameter — dedup strictness, merge strategy, recall depth, rerank mode — not a hidden model judgment.
- Every write logs what was added, merged, or skipped, so memory drift can be audited instead of discovered by accident.
Features
-
Multi-target dedup & merge
- More thorough updates at no extra call cost: when one new fact affects several existing memories, all of them get updated in one pass — not just the closest match — leaving no stale or contradictory memory behind.
- Detection mode, update behavior, and dedup itself are all switchable — see the configuration reference.
-
Less memory pollution
- Avoids saving AI suggestions, guesses, or tool noise as user facts.
- Tracks whether each memory came from the user, assistant, or tool output.
-
More relevant recall
- Multi-signal retrieval: dense vector search + BM25 keyword matching, with optional entity boosting.
- Rerank filters and reorders candidates before injection into agent context — LLM (listwise/pointwise) or cross-encoder (hosted API or local model).
-
Lightweight local storage
- Runs with local Qdrant (embedded or server mode) by default.
- Does not require Redis, a hosted memory service, or a full database stack.
-
Optional graph memory
- Entity-relation storage via KuzuDB. Off by default.
-
Agent integrations
- Works with OpenClaw, Hermes, and Claude Code.
- Python client API shaped like mem0's — point your existing mem0 client at the local server to migrate.
How it works
Add flow
messages
↓
retrieve last-k messages as extraction context
↓
LLM memory extraction (with last-k context)
↓
context completion and source tracking
↓
sequential LLM-assisted memory decisions
├─ add -> store as new memory
├─ none -> skip (duplicate)
└─ update -> merge per DEDUP_RESOLVER (skip/replace/rewrite/edit)
↓
write to vector store + BM25 index + entity store
Search flow
query
↓
dense vector search + BM25 sparse search + entity boosting
↓
rerank (LLM listwise/pointwise or cross-encoder)
↓
threshold filtering
↓
results
Compatibility
NeatMem implements a mem0-compatible API subset for local agent memory workflows:
- add memory
- search memory
- list memories
- update memory
- delete memory
- health check
It is designed to work with OpenClaw's and Hermes' memory plugin flows and other mem0-style integrations. It does not aim to cover every mem0 SDK feature or hosted-platform behavior.
Runs on 10 LLM providers and 4 embedding providers (MiniMax, DeepSeek, Qwen, GLM, Kimi, Doubao, SiliconFlow, OpenAI, Gemini, OpenRouter) — endpoints and thinking-control parameters in supported providers.
A remote client is provided for programmatic access:
from neatmem import MemoryClient
client = MemoryClient(host="http://localhost:8790") # requires `neatmem serve`
added = client.add("My name is Alex", user_id="default_user")
# {"results": [{"id": "...", "memory": "User's name is Alex", "event": "ADD"}]}
found = client.search("What is my name?", filters={"user_id": "default_user"})
print(found["results"][0]["memory"]) # -> "User's name is Alex"
Full method and parameter reference: Python Client.
The client also provides server-side write batching (add_messages, get_next_batch, mark_batch_processed, flush_messages) and raw message history access (client.messages — query, sessions, delete, reset).
Quick start
pip install neatmem
# Minimal .env (OpenAI-compatible LLM + SiliconFlow embedding)
curl -o .env https://raw.githubusercontent.com/kanhaoning/NeatMem/main/.env.example
neatmem serve # listens on http://localhost:8790
For better BM25 keyword matching (searching "memory" also matches "memories"): pip install "neatmem[nlp]" && python -m spacy download en_core_web_sm. For source install and more, see the full quick start.
Configuration
NeatMem reads configuration from environment variables (a .env file in the working directory). Common settings — full table in the configuration reference:
| Variable | Required | Default | Description |
|---|---|---|---|
LLM_PROVIDER | no | - | LLM provider preset (minimax, deepseek, dashscope, …) — supplies the default base URL |
LLM_API_KEY | yes | - | LLM API key (OPENAI_API_KEY accepted as fallback) |
LLM_MODEL | yes | - | LLM model name (no default; server refuses to boot without it) |
EMBEDDER_PROVIDER | no | siliconflow | siliconflow, openai, dashscope, or xinference |
EMBEDDER_API_KEY | conditional | - | Required for hosted embedding providers |
EMBEDDER_MODEL | no | BAAI/bge-m3 | Embedding model name |
NEATMEM_PORT | no | 8790 | Server port |
DEDUP_ENABLED | no | true | Enable dedup on write |
DEDUP_RESOLVER | no | rewrite | Duplicate resolution: skip, replace, rewrite, edit |
Custom prompts
Every core prompt (extraction, dedup, merge rewrite, group rewrite, patch edit, rerank) can be replaced with your own prompt file — see the custom prompts guide.
OpenClaw integration
With the NeatMem server running at http://localhost:8790:
openclaw plugins install @neatmem/openclaw-neatmem
openclaw neatmem init
Then restart the gateway (openclaw gateway restart) to load the plugin.
init works with zero flags: it writes apiKey=neatmem-local, baseUrl=http://localhost:8790, and your OS username as userId, then validates against the server. Override with --api-key, --user-id, or --base-url.
Example OpenClaw configuration:
{
"plugins": {
"slots": {
"memory": "openclaw-neatmem"
},
"entries": {
"openclaw-neatmem": {
"enabled": true,
"config": {
"apiKey": "neatmem-local",
"userId": "default_user",
"baseUrl": "http://localhost:8790"
}
}
}
}
}
Then check:
openclaw neatmem status
The plugin id is openclaw-neatmem. It talks to NeatMem through the local mem0-compatible HTTP API. For full CLI/tool reference and building from source, see openclaw/README.md.
Hermes integration
NeatMem includes a Hermes Agent memory provider under hermes/. With the NeatMem server running at http://localhost:8790:
hermes plugins install kanhaoning/NeatMem/hermes --enable
hermes config set memory.provider neatmem
The plugin registers four memory tools (neatmem_search, neatmem_list, neatmem_update, neatmem_delete) and recalls memories automatically on each turn. Each turn is forwarded to the server, which extracts memories in fixed-size batches; anything still pending is saved automatically when the session ends. Optional configuration via ~/.hermes/neatmem.json:
{
"base_url": "http://localhost:8790",
"user_id": "myname",
"rerank": true
}
Verify: tell Hermes "remember that I prefer dark themes", then ask about it in a new session (pending messages are saved on session switch; extraction takes a few seconds). See hermes/README.md for the full configuration reference and troubleshooting.
Claude Code integration
With the NeatMem server running at http://localhost:8790:
claude plugin marketplace add kanhaoning/NeatMem
claude plugin install neatmem@neatmem
Open a new session after installing — plugins load at session start. Sessions are captured automatically and extracted when the session ends; the first message of every new session searches and injects relevant memories. Defaults need no configuration (localhost:8790, your OS account as the memory user).
Verify: say "remember that I prefer dark themes", /exit, then ask about it in a new session in the same directory. See claude-code/README.md for the full configuration reference, command list and troubleshooting.
DeepSeek Harness integration
NeatMem includes a DeepSeek Harness (dsh) plugin under dsh/. With the NeatMem server running at http://localhost:8790:
dsh plugin --profile web add @neatmem/dsh-neatmem
Restart dsh to load the plugin — memory is on. (Using the headless CLI or another profile instead of the web UI? Swap web for that profile's name.) Verify with dsh --profile web --dump-config (a neatmem-dsh row appears). The plugin is pure TypeScript — no native dependencies and no build approvals. It works with zero configuration (baseUrl=http://localhost:8790, userId=default); override per profile in $DSH_HOME/profiles/<name>/cordis.patch.yml:
- id: neatmem-dsh
config:
userId: myname
Each direct-user turn gets one bounded automatic recall (fail-open, injected as a source-labelled message), every finished turn is forwarded to the server's /v1/messages/ batching pipeline, and the agent gets five memory tools (memory_search, memory_list, memory_get, memory_update, memory_delete). Verified against dsh 0.1.5-rc.2. See dsh/README.md for the full configuration reference and development setup.
API reference
mem0-compatible endpoints for add, search, list, get, update, delete, and health check, plus a /v1/messages/ endpoint family for server-side write batching — with curl examples in the API reference.
Roadmap
- Bilingual multi-signal support (improved Chinese/English BM25 and entity extraction)
- Memory inspection and export/import tools
- Richer recall diagnostics
License
MIT License.
Acknowledgements
Inspired by the mem0 project (Apache-2.0). Vendored-code notices are in the respective file headers.