Three weeks ago we published a competitive analysis comparing MentisDB to Mem0, Graphiti/Zep, Letta, Neo4j LLM Graph Builder, and Cognee. That analysis was accurate as of April 10. This is a refresh — same competitors, but updated to reflect what shipped in 0.8.7 and 0.8.8, and what we've learned since.
MentisDB 0.8.8 now has:
The comparison table has changed significantly. Here's what the competitive landscape looks like as of April 13, 2026.
| Feature | MentisDB | Mem0 | Graphiti/Zep | Letta | Cognee |
|---|---|---|---|---|---|
| Language / Runtime | Rust (static binary) | Python | Python | Python/TS | Python |
| Storage | Embedded (sled) | External DB | External DB | External DB | External DB |
| LLM Required for Core | No (opt-in reranking only) | Yes | Yes | Yes | Yes |
| Cryptographic Integrity | Hash chain (SHA-256) | No | No | No | No |
| Hybrid Retrieval | BM25+vec+graph+RRF | vec+keyword | semantic+kw+graph | No | vec+graph |
| Memory Branching | BranchesFrom (0.8.6) | No | No | No | No |
| Temporal Facts | valid_at/invalid_at (0.8.2) | Updates | valid_at / invalid_at | No | No |
| Memory Dedup | Jaccard threshold (0.8.2) | LLM-based | Merge | No | Partial |
| Entity Types (Custom Ontology) | entity_type field (0.8.7) | No | Pydantic models | No | Yes |
| Episode Provenance | source_episode (0.8.8) | No | Episodes | No | Partial |
| LLM Reranking | Opt-in (0.8.8) | No | No | No | No |
| RRF Reranking | RRF k=60 (0.8.6) | No | No | No | No |
| MCP Server | Built-in (0.8.0) | No | Yes | No | Partial |
| Versioned Skill Registry | Ed25519-signed (0.8.4) | No | No | No | No |
| Agent Registry + Aliases | Ed25519 keys + status | No | No | Yes | No |
| Cross-Chain Graph Queries | BranchesFrom traversal (0.8.6) | No | No | No | No |
| Webhook Notifications | Planned | No | No | No | No |
| LangChain/LlamaIndex | Planned | Yes | Partial | Yes | Yes |
| Token Tracking | Planned (1.0) | No | No | Yes | No |
| Browser Extension | Planned (1.0) | Yes | No | No | No |
new = shipped since April 10 updated = was partial/incorrect in prior analysis
Three weeks is a long time in a fast-moving space. Here's what was wrong or incomplete in our prior analysis:
The April 10 analysis listed "Temporal Fact Management" as our #1 gap and
explained how we'd implement valid_at/invalid_at on
relations. We already shipped this in 0.8.2 (April 11). The feature was on
the prior roadmap and implemented before the analysis was published — we
simply didn't update the doc in time. Temporal facts are a
solved problem for MentisDB.
Same situation. Jaccard-threshold dedup with auto-Supersedes relations was implemented in 0.8.2. The competitive analysis called this a gap; it wasn't.
The analysis said we'd need to add entity_type and
relation_type fields. We shipped the entity_type field plus a
full per-chain registry with auto-observation and persistence in 0.8.7.
The analysis recommended adding a source_episode field. We
implemented it in 0.8.8, complete with ThoughtQuery filter, dashboard
display, and JSON serialization.
This was not on our radar three weeks ago. The LongMemEval benchmark showed our lexical+vector+graph pipeline leaves room for improvement on complex multi-hop queries. LLM reranking (0.8.8) is our first explicit step toward closing that gap — opt-in, with graceful fallback if the LLM API is unavailable.
After shipping everything above, here are the honest gaps still on the 0.9.0 list:
Token tracking and the browser extension are 1.0 items — we're not rushing those.
Retrieval quality metrics as of April 13:
| Benchmark | Metric | Score | Notes |
|---|---|---|---|
| LoCoMo 10-persona | R@10 | 73.0% | Fresh chain, rebuilt vector sidecar |
| LoCoMo 10-persona w/ RRF | R@10 | 73.0% | Multi-type +0.5%; RRF neutral on simple queries |
| LongMemEval | R@5 | 57.6% | First baseline established (0.8.6) |
| LongMemEval | R@10 | 62.6% | Room to improve; LLM reranking is the path |
| Write latency | vs pre-0.8.0 | -13.8% | After 0.8.0 write performance improvements |
The LongMemEval numbers (57.6% R@5 / 62.6% R@10) are the focus area. Our hybrid BM25+vector+graph pipeline works well for single-hop factual recall (LoCoMo 73%) but gaps appear on multi-entity, multi-hop reasoning tasks. LLM reranking is the current approach — we're tuning the prompt and model selection.
Three weeks ago we called out five gaps. Two were already shipped (temporal facts, dedup). Two are now shipped (entity_type, source_episode). LLM reranking is new. The remaining roadmap is narrower: webhooks, Python bindings, and opt-in LLM extraction.
The architectural advantages we lead with — Rust, embedded storage, no-LLM core, hash chain integrity, versioned skill registry, MCP server, agent identity — remain intact and differentiated. What we've added in retrieval quality, temporal reasoning, and semantic organization closes most of the feature gap that made other systems look compelling.
MentisDB is now the only local-first, cryptographically-verified, hybrid-retrieval memory system with a built-in MCP server, versioned skill registry, and zero mandatory external dependencies — in Rust.