Durable memory for AI agents
The binary drops its trailing d — mentisdb is now
the single canonical name. The dashboard gains a Settings tab for live
env-var editing, per-chain storage size visibility, and .env
file support for cleaner project-local configuration.
Edit skills directly from the web dashboard, save changes as immutable versions, edit the version currently being viewed, and render skill Markdown links safely. Stdio mode also gets stricter MCP-client compatibility and headless background daemon launch.
Server-Sent Events streaming on MCP servers with live JSON-RPC event broadcast,
a bind_host parameter for multi-homed deployments, and an
agent_id filter on mentisdb_recent_context so each agent
loads only its own recent thoughts on shared chains.
MCP protocol upgraded to 2025-11-25 with conditional Origin validation. Hermes by Nous Research ships a native MentisDB MemoryProvider — the first third-party agent to integrate directly without an MCP bridge. Dashboard Bootstrap and Refresh buttons restored after a refactor dropped their click handlers.
Ranked search queries drop from ~35 ms to ~237 µs (99.3% faster) thanks to an
incremental LexicalIndex that is built once at chain open and updated on every
append. The dashboard Agents page now renders instantly via two-phase parallel
loading. And when MENTISDB_THOUGHT_SOUNDS=true, every read operation
plays a unique sine-wave chime — instinctively distinct from the square-wave
write sounds.
OpenClaw is a TypeScript multi-channel AI gateway with 17,000+ commits and active daily development. We study its capability-registration memory system, show the zero-code MCP integration (four lines of YAML), provide a full native TypeScript backend implementation, and compare it against Hermes to test and revise the agent-memory-protocol proposal from last week.
Hermes (Nous Research) has the most complete open-source agent memory abstraction we've
seen: a MemoryProvider ABC with turn-sync, semantic prefetch, pre-compression
extraction, built-in write bridging, delegation observability, and session lifecycle hooks.
We propose it as the basis of a community standard, show a full MentisDB implementation,
and include a zero-code tutorial for connecting MentisDB to Hermes as an MCP tool.
A full ratatui terminal dashboard for mentisdb: live log streaming,
tabbed chain/agent/skill registry, agent primer panel, and drag-to-select clipboard
copy of any text on screen via arboard and OSC 52. Plus: seamless single-TUI
lifecycle (no startup flash), startup crash overlay, log auto-scroll fix, and the
ratatui post-draw buffer swap bug that was silently blocking all text extraction.
Stdio MCP now auto-detects or launches the daemon so Claude Desktop works with zero
pre-flight. start_servers shares one MentisDbService across
every surface, ending cross-surface split brain between HTTP, HTTPS, MCP, REST, and
dashboard. Benchmarks are reproducible bit-identical: LoCoMo-10P R@10 71.9%,
LongMemEval R@5/R@10/R@20 66.8% / 72.2% / 78.0%. Plus backup/restore, admin flush,
and search correctness fixes.
A full code-level walk through mentisdb_ranked_search: filter-first
selection, as-of temporal filter, three parallel scorers (BM25, dense vector,
graph BFS), smooth exponential fusion, session cohesion, Reciprocal Rank
Fusion, and deterministic sort — every phase with file and line references.
One call, no refinement loop.
The stdio MCP process now auto-detects a running daemon, proxies to it if found, or launches one in the background. Claude Desktop users get shared live state with zero configuration — no Node.js, no mcp-remote, no TLS config.
Complete tutorial for MentisDB's LLM-extracted memories pipeline. Covers setup, REST, MCP, and Rust usage, live-tested behavior, review-before-append workflow, and real use cases for normal people, finance teams, defense workflows, and coders.
We audited the live MentisDB project chain and found a strong overconcentration in
Summary, LessonLearned, and generic References
links. This post covers the histogram, what guidance was wrong, and what we changed in
MENTISDB_SKILL.md for the next release.
A fair, detailed comparison of MentisDB and Cognee across architecture, ingestion, retrieval, provenance, operational complexity, enterprise fit, and multi-agent workflows, ending with a comprehensive side-by-side table.
Complete guide to integrating MentisDB with any MCP client. Covers Claude Desktop (mcp-remote with Homebrew), OpenCode, custom MCP clients, and the REST API for any HTTP-capable language. Includes the one-command wizard setup.
Full guide to pymentisdb — the official Python client for MentisDB. Covers the MentisDbClient for appending thoughts and ranked search, LangChain integration with MentisDbMemory, typed relations, and a complete working example.
Four days after our April 10 competitive analysis, 11 releases later: 74.0% R@10 on the full 10-persona LoCoMo benchmark (1977 queries), every feature gap closed, federated cross-chain search, pymentisdb on PyPI, and a complete competitive update covering Hindsight, Cognee, LangMem, and more.
Webhook delivery for thought append events (async HTTP callbacks with retry), irregular verb lemma expansion in lexical search, and LoCoMo stable at 72.0% R@10 with LongMemEval R@5 at 57.6%.
Optional LLM-based reranking for ranked search, source_episode field for grouping thoughts by conversational context, cross-chain relation traversal fix, and 14 new search regression tests.
Updated feature comparison against Mem0, Graphiti/Zep, Letta, and Cognee. Temporal facts and dedup were already shipped before the original analysis. entity_type, source_episode, and LLM reranking are all new since April 10.
Structured memory categories via the new entity_type field, per-chain type registry with persistence, and a full dashboard modal UX overhaul. Also fixes the wizard setup crash with older Node versions.
Complete guide to installing and running the mentisdb daemon as a proper background service using systemd. Includes dedicated user, environment configuration, security hardening, and useful management commands.
Reciprocal Rank Fusion (RRF) reranking, memory chain branching with BranchesFrom relations, irregular verb lemma expansion, and a BM25 DF cutoff fix. LoCoMo stable at 73.0% R@10. First LongMemEval baseline: 57.6% R@5.
Tuned session cohesion (radius 8→12, boost 0.8→1.2) and doubled graph relation scores. LoCoMo 10-persona R@10 jumps from 72.8% to 74.6%, clearing the 74.2% baseline. Single-hop recall hits 79.0%.
Three releases in 72 hours. A migration crash, a catch-22 where the auto-updater couldn't run because the bug blocked startup, and a runtime-in-runtime panic in the fix. The full story of what broke, why, and how we recovered.
Users upgrading from 0.8.1 hit a hard crash on chain open. The V2→V3 migration code peeked only the first thought's schema version and applied that format to the entire chain. Fixed with per-thought schema detection. 11 new migration tests. WHITEPAPER.md rewritten.
Four features closing the competitive gap: temporal edge validity for point-in-time queries, Jaccard-based auto-deduplication, User/Session/Agent memory scopes, and add/search/agents CLI subcommands on the existing mentisdb binary. No LLM calls, no cloud dependencies, no schema-breaking surprises.
Updated competitive analysis covering Hindsight (SOTA, independently verified), Cognee v1.0 with Hermes integration, LangMem ecosystem advantage, and the updated feature comparison across 6 systems. MentisDB: 74.0% R@10 LoCoMo, 15 gaps closed since April 10.
Session cohesion scoring, smooth exponential vector-lexical fusion, and a tighter BM25 DF cutoff push LongMemEval R@5 to 67.6% and LoCoMo R@10 to 88.7% — within 0.2% of MemPalace's published hybrid score.
The biggest search quality release ever. LongMemEval R@5 went from 57.2% to 65.0% with Porter stemming, tiered vector-lexical fusion, and importance weighting. Plus 13.8% faster writes, security hardening, and a skill file agents actually read.
We benchmarked MentisDB on LongMemEval — the standard benchmark for long-term memory retrieval. Starting at 57.2% R@5, we reached 65.0% through Porter stemming, tiered vector-lexical fusion, and importance-weighted scoring. Here's what worked, what didn't, and how we compare to MemX and Mem0.