Chat Memory
Conversation memory. Blind to business data, blind to events, blind to what the user did yesterday.
Research
mema, verifiable memory for AI agents.
Explicit memory records, provenance, and policy decisions outside the model. Versioned, inspectable, and explicit about limitations.
§01 what mema is
Explicit records instead of hidden model state.
mema keeps episodes, facts, and beliefs outside the model in a human-readable markdown vault. Records can be wrapped with a UAL and hash metadata as checkable assets. The core needs neither a graph database nor a blockchain; external models and anchor targets remain separate, explicit operating choices.
§02 the problem
Why agents can fail in operation.
Even a strong model fails when relevant context is missing, stale, or cannot be checked.
Static RAG
Stale and incomplete. Doesn't reflect what just happened. Doesn't track how facts change.
No Audit Trail
Without recorded sources, selection reasons, and policy decisions, an answer is difficult to trace or reproduce later.
Scattered context makes traceability and control harder.
§03 architecture
Seven layers. One vault.
Sources
ChatDocumentsTool calls
The 7-layer vault
L1Episodic
L2Semantic
L3Cognitive
L4Governance
L5Retrieval
L6Audit
L7Asset
Checkable record
MEMA RECEIPT
- kind
- fact
- hash
- sha256:ab4f…d11e
- ual
- mema://owner/01KR…
- integrity
- hash-check
CHECKABLE
Seven composable layers, from the raw event to a checkable asset record. Each has a clear purpose and an endpoint.
L1 Episodic raw events
Raw events: conversations, documents, tool calls, observations.
A write creates the external evidence record; later changes follow explicit lifecycle or tombstone paths. An episode records the source, not the truth of every derived claim.
POST /v2/observe L2 Semantic entities & facts
Entities + facts with bi-temporal validity.
valid_from / valid_to record asserted world validity; invalidated_at / superseded_by record when MEMA invalidates or replaces a claim.
POST /v2/fact L3 Cognitive beliefs & supersession
Experiences, observations, beliefs, with confidence and supersession.
The public /v2/observe path calls an LLM extractor by default and stores filtered results directly as approved on the automatic path. That is external record processing, not model-weight training. Rule-based reflection is separate; LLM reflection is opt-in.
POST /v2/cognitive L4 Governance purpose & retention
Purpose, retention, provenance, hard erasure.
Governance metadata can carry purpose, retention, jurisdiction, and evidence; policyCheck() decides at recall time. hardErase tombstones the targeted source record and preserves the audit reference. Cascade to derived records remains an explicit limitation.
POST /v2/erase L5 Retrieval search pipeline
Keyword + vector + graph + temporal + policy in one pipeline.
Fused scoring and graph expansion via derived_from return retrieval and audit metadata. In the internal v2.24+ baseline, sources carry a source class; marketing claims are demoted in ranking and stay labeled. That change is not yet in public 2.22.13.
POST /v2/recall L6 Audit hash chain
SHA-256 hash chain with a separate witness log.
verifyChain() checks ordering and hash consistency within the implemented chain/witness model. That is tamper evidence, not an independent trust anchor.
GET /v2/audit/verify L7 Asset verifiable asset
UAL + content_hash + metadata_hash + anchor lifecycle.
Records can be wrapped with a UAL plus content and metadata hashes. A local anchor receipt is publicly implemented; independent evidence export remains planned.
POST /v2/asset/wrap §04 evidence on selected retrieval paths
Retrieval metadata can serve as a receipt.
Example receipt. Available fields depend on record type, governance block, and asset wrapping.
score_componentskeyword + vector + graph fused, every component inspectable.ualstable identifier for wrapped assets. A hash comparison checks integrity, not the truth of the content.governancepolicy decision with reason. Denials on this path are logged.verification_statusunverified → verified → anchored lifecycle.
§05 the numbers
Measured in a narrow benchmark.
M01 96.0% Precision@1 v2.0 retrieval benchmark · 25 queries · 347 docs
M02 <50ms median recall latency single-shot retrieval, no agentic loop
M03 2.22.13 public source baseline Package at c37dbf7 · runtime checked 2026-08-26
M04 v2.26 internally verified build d542fba · 25 commits ahead of public main · not public
METHOD The 96.0% figure comes from an internal v2.0 retrieval benchmark with 25 keyword-anchored queries on a 347-document corpus; it is not a cross-system benchmark. A reproduced runtime run at public 2.22.13/c37dbf7 produced 536 tests across 61 files with 1,361 expectations. The README figures 538/63/1,362 are a stale documentation claim. The immutable internal v2.26 artifact records 562 tests. Those evidence baselines are not blended into a public-release claim.
§06 where mema sits
Not the same category. So no checkbox table.
A managed platform, open-source framework, local control plane, and governance overlay solve different boundaries. We compare claims documented by primary sources when checked on 2026-08-25.
Zep / Graphiti
Managed agent memory / self-operated context-graph frameworkZep now documents a proprietary managed Context Graph Engine with temporal history, provenance, RBAC/ABAC, audit, retention, and multi-tenant isolation. Graphiti is the separate Apache-2.0 open-source core, runs locally, and supports pluggable graph backends.
Cloud, BYOK, and BYOC are documented; BYOC means the customer's VPC, not a publicly documented air-gapped or customer-datacenter offer. Deletion semantics vary by scope. Benchmark figures remain vendor-reported.
SuperLocalMemory (Qualixar)
Open-source memory control plane with a local coreQualixar v4.1.9 documents SQLite + sqlite-vec as its canonical local store, with multi-channel retrieval, profile/scope isolation, and policy, provenance, audit, export, and erasure surfaces. In Mode A, core operations can run without a cloud model provider; optional downloads, connectors, backup, and other integrations remain separate network paths.
AGPL-3.0-or-later. The project says benchmark evidence carried into v4 comes from the v3 research architecture, not a fresh v4 package run. SLM-Mesh coordinates peers; it is not replicated memory. Claims remain project-reported.
Provem
Open-source research project for memory governanceProvem can use a built-in Naive/BM25/SQLite backend or wrap another store. It gates writes and recall for scope, erasure, provenance, purpose, and abstention; when wrapping a store, the project says it does not improve that store's recall.
Self-hosted, with no documented external security audit or production certification; authentication and TLS remain deployment responsibilities, physical deletion is not universally guaranteed, and benchmarks are project-authored. Published PyPI package 0.1.0 predates the cited f6ce1b6 repository revision and carries a different license.
MEMA
Explicit external memory system with evidence and policy layersPublic 2.22.13 contains the seven-layer core. Internally verified runtime d542fba is 25 commits ahead but is not public. The rendered whitepaper page is an editorial derivative, not the immutable internal artifact. MEMA changes records and retrieval, not the answer model's weights.
Production readiness is not asserted. The customer-datacenter package, independent evidence export, and HOPE integration remain clearly labelled roadmap work.
“Not publicly documented” is not evidence of absence. We therefore avoid binary no-cells when a primary source does not establish a capability.
§07 target scenarios
Three deployment pictures, not customer-reference claims.
Regulated assistant memory
Target state: a Swiss private bank operates an internal assistant for relationship managers.
MEMA is intended to record retrieval events and support targeted tombstones. Operating approval, derived-record erasure cascade, and independent evidence export must be established separately for the deployment.
Tool-call provenance
Target state: integrated tool calls are recorded as L1 episodes with source metadata.
Where the integration writes those records and derived_from links, an auditor can trace the sources used. Completeness depends on the connected workflow.
Compliance erase
Pharma R&D feeds protocols and decisions into mema.
A targeted hardErase overwrites the source record with a tombstone while preserving the audit reference. This can support part of an erasure process, but it does not replace derived-record cascade or legal and operational validation.
§08 evidence without release blending
Two source baselines, kept separate.
Public: 2.22.13 at c37dbf7.
A reproduced runtime run at the publicly inspectable baseline produced 536 tests across 61 files with 1,361 expectations. The README's 538/63/1,362 claim is stale. This is not a production approval.
Internal: v2.26 at d542fba.
The locally verified internal runtime is 25 commits ahead of c37dbf7. The immutable artifact and editorial website presentation are identified separately; neither internal evidence source is publicly available.
Benchmarks have scope.
The 25-query retrieval benchmark belongs to v2.0. Trust scenarios and test counts in the immutable artifact belong to the internal v2.26 evidence baseline, not to the editorial website presentation and not to a public release.
Production readiness: not asserted.
A release requires source parity plus current independent security, operations, backup, and recovery evidence.
mema is our ongoing, source-available research and a reference for our context engineering. Public source, internal evidence, and roadmap remain separate. Review evidence status → · Read the engineering notes →
editorial whitepaper web edition
The rendered text is a current editorial derivative, not a byte-identical
copy of the immutable internal artifact at d542fba:docs/WHITEPAPER.md.
Runtime d542fba is 25
commits ahead of c37dbf7; neither that runtime nor its source artifact is public.
The reproduced public runtime baseline is 536 tests across 61 files with 1,361 expectations. The README's 538/63/1,362 claim is stale.
Read the editorial web edition · Review the immutable artifact record