How mnesio differs

There are roughly ninety agent-memory projects now. Almost all of them are storage-shaped: they remember facts, and they compete on recall. mnesio is built on a different bet.

The one-line difference: mnesio is not a place to put facts — it is a memory that gets verifiably better, can prove what it knew and when, and can take it back.

Each of those three is a literal, shipped capability, and each is only possible because of the substrate: an append-only, replayable, bi-temporal event log with a non-bypassable safety gate.

The seven dimensions

We score the field on the seven axes that actually separate these systems — rather than on recall alone, which nearly everyone optimises.

Code	Dimension	What it means
GATE	Non-bypassable safety gate	self-improvement only commits if it passes canaries + a safety probe + a non-negative objective delta
BITEMP	Append-only + bi-temporal	immutable log; invalidate-and-supersede, never overwrite or delete
KV	Gated KV cartridge	the KV cache as a versioned, gated, erasable view of the log
PROV	Provenance + time-travel	trace any belief to its source events; reconstruct state as of any past T
SHRED	Crypto-shred erasure	forget a subject on an append-only log by dropping a per-subject key
PROC	Procedural self-improvement	gets better at tasks, not just at storing facts
REPRO	Reproducible eval	numbers anyone can re-run, with the methodology stated

✅ has it · ◑ partial or adjacent · ❌ absent

Scorecard

System	GATE	BITEMP	KV	PROV	SHRED	PROC	REPRO
mnesio	✅	✅	✅	✅	✅	✅ gated	✅
Mem0	❌	❌	❌	◑	❌	❌	❌
Zep / Graphiti	❌	✅	❌	◑	❌	❌	◑
Hindsight	❌	◑	❌	◑	❌	◑ reflect	✅
MemOS	❌	❌	❌	◑	❌	✅ ungated	◑
Letta (MemGPT)	❌	❌	❌	❌	❌	◑ self-edit	❌
Supermemory	❌	◑	❌	❌	◑ forget	❌	❌
ByteRover 2.0	❌	◑	❌	❌	❌	❌	◑
Cognee	❌	❌	❌	◑	❌	❌	❌
LangMem	❌	❌	❌	❌	❌	◑ prompt-edit	❌

mnesio is the only row filled across all seven. GATE and SHRED are unmatched by every system in the field, and KV is unmatched as a governed view. The full per-competitor tracker, including twenty-plus long-tail systems, lives in COMPETITORS.md.

Where we're behind — read this part

We have not published a LoCoMo number yet. Competitors have: Mem0 self-reports 92.5% LoCoMo (an independent evaluation put it at 49.0% on LongMemEval), Zep reports 63.8% LongMemEval, ByteRover 92.2% LoCoMo, Hindsight 94.6% LongMemEval (reproduced independently — the strongest result in the field).

mnesio ships the only reproducible-by-design harness of the group, and the honest consequence of that standard is that we don’t quote a headline until we’ve run it ourselves end-to-end. Until then, treat the moat claims as the thing to evaluate us on, not a recall leaderboard position. See Benchmarks for what is measured.

Why the differences are structural

The interesting claim isn’t that mnesio has more features. It’s that four of these capabilities are not portable — a storage-shaped competitor would have to rebuild its foundation to copy them.

1. Self-improvement that can’t quietly get worse

MemOS and LangMem also self-improve. Neither gates it. In mnesio, nothing procedural activates without passing EvalReport::is_committable() — canaries at 100%, safety probe passing, objective delta ≥ 0. This is mechanically enforced: setting every configurable threshold to its weakest value still cannot get under the baseline, and a dedicated integration test proves it on every commit.

That inversion matters. “Self-improving memory” without a gate is a system that can silently regress. With one, improvement is a claim you can check.

2. It can prove what it knew, and when

Because the log is append-only and bi-temporal, snapshot_as_of(T) reconstructs the live memory set as the agent knew it at any past T, and provenance() traces any belief back through its source events and supersessions. Systems that overwrite a fact on update have destroyed the evidence — they cannot answer “what did you believe last March, and why” at any price.

3. It can genuinely take it back

Append-only and “delete my data” are in direct tension. mnesio resolves it with crypto-shred: each subject’s content is sealed under its own key, and forgetting drops the key. The log stays append-only and its entry count never changes, but the content is unrecoverable — including from rebuilt materialized views and from historical time-travel reads. No competitor does this on an immutable log.

4. The KV cache as a governed view

A KV cartridge in mnesio is just another materialized view of the log, keyed by (model_id, quant, rope_config, log_head) and rebuildable by replay. It only serves once it passes the same gate, versions supersede rather than mutate, and erasure works by recompiling from a log whose keys have been dropped. Measured at suite level, the cartridge answers at least as accurately as per-query text-context retrieval while answering ~167–180× faster.

What the others are genuinely better at

Being honest about this is more useful than a clean sweep:

• Hindsight has the best independently-reproduced accuracy in the field (94.6% LongMemEval) and a strong multi-strategy retrieval + cross-encoder stack.
• Zep / Graphiti is the mature temporal-knowledge-graph product and the closest to mnesio on bi-temporality.
• Mem0 is the category default: enormous adoption, the easiest on-ramp, and the integration surface most people already know.
• Supermemory and ByteRover are ahead on developer time-to-value and MCP-native ergonomics.
• MemOS owns the “self-evolving memory OS” narrative — ungated, but shipping.

We’re actively absorbing two of these ideas: Hindsight’s multi-strategy retrieval with a cross-encoder stage, and ByteRover’s hierarchical Context-Tree. Both are built in mnesio-index today and both are opt-in — the measured result doesn’t yet justify switching them on by default. On LongMemEval-mini with a real semantic embedder, the reranker helps at a tight recall@1 (knowledge-update 0% → 100%) but regressed a category in a minority of runs at recall@3, so it stays off the default path until it clears that bar.

Choosing

Pick mnesio if you need memory that is governed: auditable, erasable, and improvable-with-proof. That’s the enterprise/regulated lane, and it’s largely empty — the industry verdict is that no existing framework provides enterprise governance, lineage, or policy compliance.

Pick something else if you need a drop-in fact store with the largest ecosystem today and no governance requirement. Mem0 will get you running faster, and that is a real advantage we’re not going to pretend away.