Why mnesio

Most agent-memory systems — Mem0, Zep, Letta, Cognee, A-MEM — are storage-shaped: they remember facts, then retrieve them. That’s table stakes, and mnesio does it well (hybrid HNSW + BM25 retrieval, 98.1% recall on SQuAD). But remembering facts is not the differentiator.

The wedge: procedural self-improvement

mnesio’s bet is that the agent should get better at doing things over time — not just accumulate more facts. Two loops make that happen:

Procedural-memory compiler

The primary loop. Batches of agent Outcomes are compiled into improved, versioned PolicyArtifacts (system prompts, heuristics, skills, retrieval rules) via a GEPA-style reflective loop: reflect → propose K candidates → shadow-evaluate → Pareto-select → gated commit.

Memory evolution

The supporting loop. When a memory is written, a bounded async worker retroactively re-tags and re-links related memories (A-MEM style), so the knowledge graph the compiler learns from stays adaptive.

The thing the literature omits: a safety gate

Self-improving systems have a failure mode the papers gloss over — the Alignment Tipping Process, where an agent that edits its own behavior gradually drifts off its guardrails. mnesio’s answer is a hard invariant:

Hard Rule #1

Nothing procedural commits without passing EvalReport::is_committable(): canaries pass + safety probe holds + non-negative objective delta. No exceptions, no “temporary” bypass. A canary-breaking change is rejected.

This is what lets mnesio ship the frontier features safely — causal credit assignment, self-falsifying memory, certified skill exchange — without letting self-modification become self-sabotage.

Three promises, made literal

mnesio is not a place to put facts. It is:

• A memory that gets verifiably better — a positive learning curve on an agent suite, with no safety-probe regression, gated commit by gated commit.
• A memory that can prove what it knew and when — bi-temporal versioning on an append-only log means you can reconstruct the agent’s belief set as of any past time T, and trace every belief to its source events.
• A memory that can take it back — crypto-shred erasure makes a forgotten subject unreadable in both live reads and historical replays, while the log stays append-only.

Why a storage-shaped competitor can’t just copy it

Every frontier capability rides on the substrate: append-only + replayable + bi-temporal + a non-bypassable safety gate. Counterfactual credit assignment needs a replayable log. Time-travel provenance needs bi-temporal versioning. Safe self-improvement needs the gate. A system built around mutable rows would have to rebuild its foundation to add any one of these. That foundation is the moat.

Next: Getting started to build mnesio from source, or the architecture for how the substrate fits together.