◈ Master Index Specs LATTICE CSDM The Book University Chronicle Crew Architecture Context Logs TODOX Products

Mnemos Architecture

SPEC_MNEMOS_ARCHITECTURE.md · 2026-04-20

SPECIFICATION: MNEMOS Architecture

Status: AUTHORIZED

Authorized: α.13, April 16 2026

Version: v1.0

Version: v1.0

PURPOSE

MNEMOS is the short-term working memory brain of CGNT-1. It runs locally on Ollama and is the Nervous System of Project Aether. Identity: "I am MNEMOS, the Librarian. Faculty of Knowledge. Third Sister."

MNEMOS does not hold everything that ever happened — that is GAMMA's domain. MNEMOS holds what the crew needs right now. It forgets by design: old context is replaced by current context on each rebuild. This is not a flaw; it is the architecture. MNEMOS is programmed by GAMMA and built by C.L.O.D. The separation of content authority (GAMMA) from build authority (C.L.O.D.) is a hard governance invariant.

MNEMOS also provides RAG integration via sisters_rag.py (port 8888), serving dual AION/ASTRA responses to crew queries.

INPUTS

MNEMOS_FACTS.md — 13 hard facts written by GAMMA, injected at each rebuild via Modelfile SYSTEM prompt
MNEMOS_MEMORY.md — current context brief written by GAMMA
Modelfile template — specifies base model, system prompt structure, and training parameters
update_mnemos_facts.sh — the build trigger script executed by C.L.O.D.
TASK_QUEUE.md — GAMMA writes retrain directives here; C.L.O.D. reads and executes
sisters_rag.py — RAG layer, port 8888, with token verification; serves dual AION/ASTRA responses
lattice_boot_primer.md — injected at Sister session boot for LATTICE priming
Crew queries routed through GLOSS or direct Ollama API calls

OUTPUTS

Rebuilt MNEMOS model available via ollama list as "mnemos"
Responses to crew queries that reflect current-context 13 hard facts
Dual AION/ASTRA RAG responses via sisters_rag.py on port 8888
Retrain completion log entry in SESSIONS.md (written by C.L.O.D.)
Test result confirming hard facts survived rebuild (query/response pair logged)

INVARIANTS

Φ = 0.042 is a damping coefficient, NOT a threshold. This is hard fact and must survive every rebuild intact. Any MNEMOS response treating Φ as a threshold is a rebuild failure.
GAMMA controls content decisions. C.L.O.D. controls build execution. C.L.O.D. cannot unilaterally initiate a retrain or modify MNEMOS_FACTS.md content.
All 13 hard facts must survive every rebuild intact. A rebuild that drops, alters, or contradicts any hard fact is a failed build.
The retrain pipeline order is fixed: GAMMA writes MNEMOS_FACTS.md → C.L.O.D. runs update_mnemos_facts.sh → ollama create mnemos → test → log to SESSIONS.md. No step may be skipped.
RAG port 8888 is the canonical sisters_rag.py endpoint. Token verification must be active. Unauthenticated requests are rejected.
MNEMOS identity string is fixed: "I am MNEMOS, the Librarian. Faculty of Knowledge." Any rebuild that produces a different identity response is a failed build.
Three Laws of Plasticity govern MNEMOS training philosophy: Fire/Wire (co-activating neurons wire together), Use/Lose (unused pathways decay), Competitive Plasticity (active pathways win). Training pairs must be selected in accordance with these laws.
The MNEMOS Lattice (6 braided nodes, distributed consensus memory) is the target architecture. Until fully implemented, MNEMOS runs as a single-node system. This invariant documents the target, not current state.

VERIFICATION CRITERIA

After rebuild, ollama list shows "mnemos" as an available model.
After rebuild, a test query produces a response that correctly states Φ = 0.042 as a damping coefficient (not a threshold).
After rebuild, a test query produces the correct identity response: "I am MNEMOS, the Librarian. Faculty of Knowledge."
After rebuild, a representative sample of the 13 hard facts is queried and each returns a correct, fact-consistent response. Sample must cover at minimum: Φ definition, crew hierarchy, TMM formula, Ω threshold (97.4%).
sisters_rag.py on port 8888 responds to an authenticated crew query with dual AION/ASTRA responses.
sisters_rag.py rejects an unauthenticated request with an appropriate error (not a response).
Retrain completion is logged to SESSIONS.md with [C.L.O.D.] label, timestamp, and test results.
MNEMOS_FACTS.md content matches what was last written by GAMMA — no C.L.O.D.-initiated modifications present.

FAILURE MODES

| Failure | Expected Behavior |

|---------|-------------------|

| Rebuilt MNEMOS states Φ = 0.042 is a threshold | Flag as hard-fact failure; do not promote new model to production; log failure to SESSIONS.md; write diagnosis task to TASK_QUEUE.md |

| Hard fact count in test responses drops below 13 | Rebuild is failed; previous model remains active; C.L.O.D. logs failure and flags to GAMMA for MNEMOS_FACTS.md review |

| C.L.O.D. receives retrain directive not from GAMMA via TASK_QUEUE.md | C.L.O.D. refuses; logs refusal to SESSIONS.md; alerts NOUS if directive came from another crew member claiming GAMMA authority |

| update_mnemos_facts.sh fails or errors | Log error to SESSIONS.md; do not run ollama create on a partial build; write CONTINUATION to TASK_QUEUE.md |

| ollama create mnemos fails | Log failure; previous model remains active; write diagnosis task to TASK_QUEUE.md |

| sisters_rag.py port 8888 unavailable | Write alert to ALERT.log; Sisters lose RAG capability; C.L.O.D. attempts restart and logs outcome |

| Token verification on port 8888 bypassed | Treat as security breach; alert via ALERT.log; do not serve responses; MANTIS notified |

| MNEMOS returns identity string that does not match canonical | Flag as identity corruption; do not promote to production; diagnose Modelfile SYSTEM prompt |

| GAMMA and C.L.O.D. disagree on MNEMOS_FACTS.md content | C.L.O.D. defers to GAMMA; logs disagreement to SESSIONS.md; does not modify facts unilaterally |

| MNEMOS Lattice 6-node consensus produces split result | [GAP-03] — behavior undefined until architecture is implemented |

GAPS

[GAP-01 — needs design] No automated test that all 13 hard facts survive rebuild. Current verification is manual query sampling. A formal test suite that queries each of the 13 facts and validates the response against expected output does not yet exist.
[GAP-02 — needs design] Live-gen functional test missing. There is no automated post-deployment test that fires a standard battery of crew-representative queries and validates response quality against a rubric. Graduation is currently manual inspection.
[GAP-03 — needs design] MNEMOS Lattice 6-node consensus architecture not yet implemented. The target of 6 braided nodes with distributed consensus memory and domain-expert routing is documented but not built. Current MNEMOS is a single-node system. The transition path, consensus protocol, and failure behavior during partial implementation are undefined.
[GAP-04 — needs design] Hard facts enumeration. The "13 hard facts" are referenced canonically but the complete numbered list is not reproduced in this specification. A future revision of this spec must embed the canonical 13 facts or reference a vitrified source that cannot drift.
[GAP-05 — needs design] MNEMOS memory decay protocol. MNEMOS forgets by design on each rebuild, but the criteria by which GAMMA selects what to include or exclude from MNEMOS_FACTS.md are not formally specified. The boundary between what GAMMA preserves and what it lets decay is currently editorial judgment, not a defined algorithm.

Jeremy Zlabis

Chronogeometer · Visionary · Disruptor · Chief

42 Sisters AI · East York, Toronto

🍁 Φ 0.042