Hermes separation / checkpoint 003
Memory landscape intake - 25 July 2026

Memory is a loop, not a filing cabinet.

The supplied list contains 25 links representing 24 unique systems, plus a pinned Memory-OS rulebook. We will measure whether an agent captures, retrieves, recognizes, uses, revises, links, and explains memory during real work.

Status: intake normalizedLocal clones: 18 / 24New systems checked: 6Runtime changes: noneMaintenance: not yet required
Architecture correction: one package need not do everything. A strong Nagato stack may combine a thin Hermes lifecycle adapter, a durable local store, and an independently updateable reasoning/consolidation layer.
Evaluation model

Separate the three jobs

A good database cannot pass with weak agent use, and a good injection rule is not mistaken for a complete memory system.

1 - Runtime loop

Capture and active use

Hermes lifecycle hooks, bounded injection, explicit identity, completed-turn sync, pre-compression, correction tools, and failure-soft behavior.

+
2 - Memory authority

Store and recall

Facts and episodes, provenance, temporal correction, isolation, portability, inspectability, local-first operation, and predictable latency.

+
3 - Learning layer

Link and consolidate

Discover relationships, combine evidence, propose higher-order memories, supersede stale claims, and leave human-readable receipts.

Notes converted to tests

What must be observable

This preserves what made the fabric approach reliable without hard-coding one provider's block names forever.

1
Recognize injection

The agent identifies the canonical envelope and memory IDs it used before acting.

2
Change the plan

A trace proves recall materially changed a decision; retrieving text alone fails.

3
Write deliberately

Updates carry provenance, scope, evidence type, and reason - never an opaque transcript dump.

4
Correct without amnesia

New facts supersede or qualify old ones while retaining temporal history.

5
Generate linked knowledge

Derived memories cite their sources and remain reviewable.

6
Human inspection

A dashboard, open notes, or evidence browser shows what exists and why it was recalled.

7
Isolation

Personal, Nagato, game, and candidate namespaces cannot bleed through search or maintenance.

8
Replacement

Adapters, export, backup, delete, and replay tests prevent irreversible lock-in.

Normalized landscape

All 24 systems have a lane

Disposition is test priority, not a verdict. Older findings are dated 29 June; six newcomers were checked against their current repository documentation on 25 July.

SystemShape / valueDisposition
MnemosyneHermes providerSQLitedashboard Known-good lifecycle baseline.Run first - control
Engraphislocal enginegraph Inspectable recall, temporal history, proactive packets, consolidation, and code links.Run first
SMRITI (vn-envy)SQLitetemporal Episodes plus facts, supersession, multi-channel retrieval, offline tests.Run first
mindone-filedream cycle Deterministic graph, provenance, aging, consolidation receipts.Run first
Open Second BrainHermes provideropen notes Human-readable control with a heavier Bun/Node surface.Run first
Honchoservicereasoning Direct reference for the desired theory-of-mind direction.Run first
ctxMCPDream Mode Autonomous cross-referencing, stale detection, temporal gravity, tenant scopes.Architecture challenger
Cogneeplatformknowledge graph Serious reasoning platform with a heavier envelope.Architecture challenger
Hindsightservice Strong retain/recall concepts; infrastructure cost is part of the test.Architecture challenger
AriadneHermes provider Promising native shape; packaging/path defects need recheck.Integration challenger
engRAMHermesencryptedoffline Bundled embeddings and encrypted RAM-indexed vault.Privacy challenger
HyAtlas MemoryHermes providergraph/vector Broad capability, heavier dependencies.Second screen
SuperLocalMemoryHermes provider License, dependency, and documentation risks.Risk screen
Memory-OSruntime modrulebook Mandatory pre-action protocol is design evidence.Pattern donor
agent-memory-skillskillgovernance Authority and provenance patterns.Pattern donor
hermes-memlockcompanion Lifecycle and safety ideas rather than a store.Adapter donor
Rhizomeknowledge vault Project knowledge rather than full conversational memory.Artifact screen
Unforgitrepo memory Useful coding context; narrower than Nagato's needs.Artifact screen
Context MCPproject context Context utility, not a full learning loop.Artifact screen
smriti (thelonewolf)young MCP Distinct from vn-envy's project.Quick screen
NoshyMCP utility Small surface; unlikely to cover the lifecycle alone.Quick screen
remembertool/library Simple surface; earlier license ambiguity.Risk screen
dyadcorecomponent Building block, not a complete provider.Component donor
edgeless-memorylibrary Basic component with little active-use evidence.Quick screen
Evaluation funnel

Broad evidence, then expensive tests

The first wave contains different architectures so we can discover which shape works.

Freeze the contract

Canonical envelopes, lifecycle adapter, rubric, privacy fixtures, corrections, negative-relevance cases, and trace receipts.

Reproducible harness

Cheap screen - all 24

License, Windows, offline posture, isolation, export/delete, activity, weight, integration, and failure behavior.

No unknown blockers

First wave - six shapes

Mnemosyne, Engraphis, vn-SMRITI, mind, Open Second Brain, and Honcho. ctx or Cognee can displace a lane on evidence.

Measured shortlist

Active-use finals - up to four stacks

Game and development scenarios test changed decisions, corrections, synthesis, irrelevant recall, restart, and restore.

Evidence-backed winner

Disposable Nagato integration

The winner runs behind the provider boundary with isolated candidate state and no live memory import.

Cutover proposal
Evidence ledger

Preserve why the links matter

These notes remain requirements even when their originating system is eliminated.

User notes

Desired product qualities

  • Mnemosyne is a better baseline than tiny flat files.
  • Memory-OS's mandatory pre-action rule tackles agents that notice memory but do not use it.
  • The fabric approach worked: base injection plus explicit update tools.
  • Open notes are attractive; Obsidian is not required.
  • Mnemosyne's dashboard is valuable.
  • Memories should combine and link through a reasoning layer; Honcho is a key reference.
Prior research

Useful, but dated

  • The 29 June report covered 18 systems and ranked Mnemosyne first.
  • Its smoke commands and current-state claims will not be reused blindly.
  • Engraphis, mind, Cognee, vn-SMRITI, ctx, and engRAM materially expand the landscape.
  • Repository claims guide tests; they do not count as our product evidence.
Human checkpoints

No maintenance window yet

Research and disposable tests do not require downtime. Live changes remain behind explicit approval.

When I will need you

I will return after the executable first wave if the evidence exposes a real product tradeoff. A maintenance window is needed only after disposable Nagato integration passes and a concrete cutover plus rollback package exists. Personal Hermes, Nagato Hermes, GTC, and the bridge will not be restarted or repointed before that checkpoint.