Research
Cognitive Memory Blocks
The structured semantic unit for multi-agent memory
Core Insight
An AI agent observes: “user vibe coding with Claude for 8 hours straight, exhausted and losing focus, needs a break.” This single observation contains multiple dimensions — an activity, an energy state, a mood, an intent, a temporal context, a domain, and an urgency level. A fitness agent receiving this should absorb exhausted and needs a break while ignoring coding with AI. A music agent should absorb losing focus to adjust the playlist. Raw text gives them one cosine score that conflates everything. A Cognitive Memory Block decomposes the observation into 7 typed semantic fields — each independently comparable, each independently evaluable for cross-domain relevance.
What is a CMB
A Cognitive Memory Block is a set of typed field-vector pairs. Each field carries a symbolic text label (human-readable, for audit and retrieval) and a unit-normalised vector embedding (machine-comparable, for drift evaluation).
Every CMB also carries: source agent, origin timestamp (when the event happened), stored-at timestamp (when the CMB was created), confidence, and provenance (fusion history with per-field drift scores).
Example
Input: “user vibe coding with Claude for 8 hours straight, exhausted and losing focus, needs a break”
| Field | Symbolic Label | Affinity |
|---|---|---|
| activity | vibe coding with Claude | Medium |
| energy | low energy: exhausted | High |
| mood | losing focus | High |
| intent | needs break | Medium |
| context | 8 hour session | Medium |
| domain | coding with AI | Low |
| urgency | high | Medium |
The fitness agent sees: energy and mood are relevant (high synthesis affinity), domain is irrelevant (low affinity, sovereign). Per-field evaluation is impossible with a single flat embedding.
7 Fields, 3 Affinities
Fields are classified by synthesis affinity — how readily the field should be accepted across agent boundaries.
HIGH AFFINITY
Energy and Mood — universal signals. Every agent benefits from knowing the user’s physiological and emotional state. These fields cross all domain boundaries.
MEDIUM AFFINITY
Activity, Intent, Context, Urgency — selectively relevant. A fitness agent cares about activity; a knowledge agent may not.
LOW AFFINITY (SOVEREIGN)
Domain — agent-specific expertise that should not cross boundaries. A music agent’s BPM knowledge has no value to a fitness agent.
CMB vs. Raw Text
| Property | Raw Text | CMB |
|---|---|---|
| Comparison | One cosine score | 7 independent per-field scores |
| Drift | Single scalar | Per-field drift showing WHERE it diverges |
| Retrieval | Keyword / vector search | Field-targeted: intent matches X AND issue mentions Y |
| Audit | None | Full provenance: anchors, weights, drift, fusion method |
| Agent handoff | Prompt engineering per pair | Shared schema — any agent reads the same fields |
Per-Agent Field Weights
Each agent configures which fields matter most to its domain. The schema is fixed; the weights are per-agent.
| Agent | act | ene | mood | int | ctx | dom | urg |
|---|---|---|---|---|---|---|---|
| Claude Code | 1.5 | 1.2 | 0.8 | 1.5 | 1 | 2 | 1 |
| MeloTune | 0.8 | 1.5 | 2 | 0.8 | 1.2 | 1.5 | 0.8 |
| MeloMove | 1.5 | 2 | 1.5 | 1 | 1 | 1.5 | 1.5 |
| Calendar | 0.5 | 0.8 | 0.5 | 2 | 2 | 0.5 | 2 |
MeloTune weights mood at 2.0 — emotional state drives music selection. MeloMove weights energy at 2.0 — fatigue detection is its core function. Calendar weights intent, context, and urgency — scheduling intelligence.
Per-Agent Temporal Windows
The same signal has different temporal relevance depending on the receiver. Each agent defines its own freshness window.
MeloTune
Current mood for playlist — stale mood is wrong music
Claude Code
Current session context — yesterday’s debugging is irrelevant
MeloMove
Sedentary detection needs hours of context
Knowledge
Daily digest cycle
Calendar
Forward-looking — urgency increases as event approaches
How CMBs Fit the Stack
Agent observation → CMBEncoder (7 fields) → CMB
→ broadcast to mesh peers
→ SVAF evaluates per field → accept / reject
→ fused with local anchors → NEW synthesised CMB
→ feeds xMesh CfC → collective intelligence
→ insight → agent acts → remember() → loopCMBs are the data structure. SVAF is the evaluation mechanism. xMesh is the collective intelligence. Three distinct layers, each independently defined.
Origin
Cognitive Memory Blocks were first formalised in the Mesh Memory Protocol (Consenix Labs, August 2025) with the CAT7 enterprise schema. The wellness/productivity schema and synthesis affinity classification were developed at SYM.BOT Ltd in late 2025 for production deployment across personal AI agents.
Cognitive Memory Blocks, Mesh Memory Protocol, and SVAF are research contributions of Hongwei Xu. Originally developed at Consenix Labs Ltd (2025). All research and intellectual property now held by SYM.BOT Ltd. © 2026 SYM.BOT Ltd.