MentisDB 0.8.2 — Temporal Memory, Auto-Dedup, Scoped Recall, CLI Subcommands

0.8.2 closes the four biggest feature gaps we identified in our competitive analysis against Mem0, Graphiti/Zep, Letta, and Cognee. Every major agentic memory system supports temporal facts, deduplication, scoped visibility, or some form of CLI — and now MentisDB does too, without adding any cloud dependencies, LLM calls, or schema-breaking migrations.

At a Glance

1. Temporal Edge Validity

Agents accumulate facts that go stale. "Alice's office is on floor 3" is true until she moves to floor 7. Competitors handle this with mutable graph properties or LLM-extracted temporal triples. MentisDB does it with append-only temporal relations.

Every ThoughtRelation now carries valid_at and invalid_at timestamps — set explicitly when you know the time bounds, or auto-set to the append time when a relation is created. The real power comes from the as_of parameter on ranked search:

let query = RankedSearchQuery::new()
    .with_text("Alice's office")
    .with_as_of("2026-03-01T00:00:00Z".parse().unwrap());

An as_of query excludes thoughts appended after the timestamp and thoughts that have been superseded, corrected, or invalidated by thoughts appended at or before it. The invalidated set is built once at chain open time and updated incrementally on each append — O(1) lookup at query time, no LLM needed.

This is a schema V3 change: ThoughtRelation expanded from 3 fields to 5. Existing V0, V1, and V2 chains are migrated automatically on open, with full hash-chain rebuild. The migration is persisted so future opens skip it.

2. Memory Deduplication

Agents in long sessions tend to re-assert facts they've already recorded. "The sky is blue" appears three times across different turns. Without dedup, all three compete for retrieval slots, wasting bandwidth and diluting relevance.

Set MENTISDB_DEDUP_THRESHOLD=0.85 (or call db.with_dedup_threshold(Some(0.85))) and every new thought is checked against the last 64 thoughts in the chain using Jaccard similarity on normalized lexical tokens. If the best match exceeds the threshold, a Supersedes relation is automatically added pointing to the prior thought.

# Enable auto-dedup
MENTISDB_DEDUP_THRESHOLD=0.85
MENTISDB_DEDUP_SCAN_WINDOW=64  # default

The superseded thought is then filtered from search results — no LLM call, no embedding comparison, no cloud dependency. Pure token-set math in microseconds.

Why Jaccard instead of embeddings? Two reasons: (1) it's deterministic and fast — no model loading, no GPU, no variance between runs; (2) for near-exact duplicates (the common case), Jaccard is more precise than cosine similarity on short texts. A threshold of 0.85 catches re-statements while allowing genuine elaborations ("The sky is blue" vs "The sky is blue on clear days" — Jaccard ≈ 0.57, below threshold).

3. Multi-Level Memory Scopes

In multi-agent systems, not every memory should be visible to every agent. A session-scoped scratchpad shouldn't pollute a user's long-term recall. An agent's private strategy notes shouldn't leak to other agents on the same chain.

MentisDB 0.8.2 introduces MemoryScope — three visibility levels stored as tags on thoughts:

Scopes are stored as tags, not as a new struct field. This was a deliberate choice: adding a field to Thought would require schema V4 and another migration round. Tags already exist, already participate in search filtering, and the scope: prefix convention is easy to audit. A helper method makes it seamless:

let input = ThoughtInput::new(ThoughtType::FactLearned, "Draft hypothesis")
    .with_scope(MemoryScope::Session);

let query = RankedSearchQuery::new()
    .with_text("hypothesis")
    .with_scope(MemoryScope::Session);

The REST API accepts "scope": "session" on both append and search. The MCP surface passes it through tags as well.

4. CLI Subcommands

The mentisdbd binary already had setup and wizard subcommands for integration configuration. 0.8.2 adds three more that talk to a running daemon via REST, giving operators a quick way to add, search, and inspect memories without opening the dashboard or writing an MCP client:

# Add a thought
mentisdbd add "The database uses schema V3" --type fact-learned --tag architecture

# Search
mentisdbd search "schema version" --limit 5 --scope user

# List agents
mentisdbd agents --chain my-project

The CLI uses ureq for synchronous HTTP — no async runtime, no Tokio dependency, no startup latency. It connects to the daemon's REST port (default http://127.0.0.1:9472) and returns JSON output.

This is especially useful for quick inspection, debugging, and scripting. Combined with jq, you can pipe search results into shell pipelines:

mentisdbd search "deployment" --limit 20 | jq '.hits[].thought.content'

Migration Safety

0.8.2 includes the third schema migration in MentisDB's history (V0 → V1 → V2 → V3). Each migration has taught us something:

• V0 → V1: added chain_key to ThoughtRelation for cross-chain references.
• V1 → V2: schema version constant introduced; migration framework established.
• V2 → V3: added valid_at/invalid_at to ThoughtRelation.

The key lesson from V3: bincode's empty-Vec fast path is a trap. When a Rust struct has Vec<T> and the serialized data has 0 items, bincode can successfully deserialize even if T's layout changed between schema versions. The deserializer reads length=0, skips the element loop, and continues at the same byte offsets — producing silently wrong data. The fix: always check schema_version == MENTISDB_CURRENT_VERSION after a successful fast-path deserialization.

Design Philosophy

All four features follow the same principles that distinguish MentisDB from the field:

1. No LLM calls at append or query time. Dedup uses Jaccard on tokens, temporal uses timestamps, scopes use tags. Every operation is deterministic and O(1) or O(scan_window) at worst.
2. Append-only integrity. Temporal invalidation is query-time only. No existing thought is ever mutated. Every historical hash remains verifiable.
3. Zero new dependencies for core features. Jaccard uses the existing lexical tokenizer. Scopes use existing tags. The CLI subcommands add ureq but it's only needed for those subcommands, not the library.
4. Backward-compatible by default. Dedup is opt-in. Scopes default to User. Temporal edges default to None. Old chains migrate transparently.

Upgrade

cargo install mentisdb

Or from source:

git pull
cargo install --path . --locked

Existing chains (V0, V1, V2) are migrated to V3 automatically on first open. The migration persists so subsequent opens are instant. Vector sidecars and skill registries are unaffected.

To enable dedup, set the environment variable before starting the daemon:

MENTISDB_DEDUP_THRESHOLD=0.85 mentisdbd

What's Next

0.8.3 will focus on retrieval quality: irregular verb lemma expansion ("went" → "go"), lightweight reranking, and per-field BM25 DF cutoffs. These address the stemming gap that accounts for ~38% of remaining LoCoMo misses. Beyond that: custom ontology support and episode provenance, the two remaining gaps from the competitive analysis.

MentisDB is an open-source durable memory layer for AI agents. It stores memories in an append-only hash-chained log, retrieves them with hybrid lexical+semantic+graph search, and runs entirely locally with no cloud dependencies. GitHub · Docs · Website