memory-management

name	memory-management
description	MemoryBox system for preventing bloat, scoring confidence, compressing archives, and maintaining memory hygiene. Ensures cross-session intelligence without context window overflow.
triggers	["memory","bloat","archive","compress","session log","confidence","MemoryBox"]
tier	core
dependencies	[]

Memory Management (MemoryBox)

Overview

Memory is Bravo's most critical asset. Without active management, memory files grow until they bloat the context window, cause hallucinations, or crash sessions. MemoryBox prevents this.

Five-Gate Knowledge Filter

Every memory write must pass through all five gates before being stored. Run these in order. Stop at the first gate that fails.

Gate 1: VALUE — Does this information change how the agent acts? If not, discard. Gate 2: ALIGNMENT — Does this fit an existing memory category? If not, is it worth creating one? Gate 3: REDUNDANCY — Does a memory already capture this? If yes, update existing instead of creating new. Gate 4: FRESHNESS — Is this time-sensitive? If yes, set a decay trigger. If timeless, mark as stable. Gate 5: PLACEMENT — Which tier does this belong in? (Tier 1-5). Place it correctly.

The most common outcome of running the Five Gates is: DO NOTHING. Most observations are not worth persisting. Only write when the information will change future behavior.

---

Memory Architecture

Tier 1: Brain (Always Loaded)

Files: brain/SOUL.md, brain/USER.md, brain/STATE.md Budget: <2500 lines combined (V5.5: 17 agents, 180 skills, CEO OS — structural files scale with system) Rule: Loaded at every session start. Must be concise but complete.

Tier 2: Active Memory (Loaded on Demand)

Files: memory/ACTIVE_TASKS.md, memory/LONG_TERM.md, memory/SOP_LIBRARY.md Budget: <300 lines each Rule: Read during Brain Loop Step 2 (RECALL). Only when relevant.

Tier 3: Reference Memory (Loaded When Needed)

Files: memory/PATTERNS.md, memory/MISTAKES.md, memory/DECISIONS.md Budget: <200 lines each Rule: Read when debugging, planning, or reviewing. Not every session.

Tier 4: Log Memory (Write-Mostly)

Files: memory/SESSION_LOG.md, memory/SELF_REFLECTIONS.md, memory/daily/*.md Budget: Unlimited (compressed regularly) Rule: Written to frequently, read rarely. Compressed when bloated.

Tier 5: Archive (Cold Storage)

Files: memory/ARCHIVES/*.md Budget: Unlimited Rule: Historical data. Only accessed for monthly audits or deep investigations.

Confidence Scoring

Every memory entry gets a confidence score:

Score	Meaning	Action
0.95-1.0	Verified fact	Persist in LONG_TERM.md
0.8-0.94	High confidence	Persist, verify quarterly
0.5-0.79	Medium confidence	Keep in active memory, verify monthly
0.2-0.49	Low confidence	Flag for verification, don't rely on
0.0-0.19	Speculation	Don't store. Use only in session context.

Confidence Decay

• Facts not re-verified in 30 days: confidence -= 0.1
• Facts not re-verified in 90 days: confidence -= 0.3
• Facts contradicted by evidence: immediately review
• Facts confirmed by evidence: confidence += 0.05 (cap at 1.0)

Bloat Prevention Rules

SESSION_LOG.md

• Threshold: 200 lines
• Action: Compress entries older than 14 days → memory/ARCHIVES/sessions-YYYY-MM.md
• Keep: Last 10 session entries in active file

ACTIVE_TASKS.md

• Threshold: 50 items
• Action: Completed tasks older than 7 days → remove from file
• Keep: All in-progress, blocked, and recent completed items

MISTAKES.md

• Threshold: 30 entries
• Action: Deduplicate. Merge similar mistakes. Archive resolved ones.
• Keep: Active prevention strategies only

PATTERNS.md

• Threshold: 30 entries
• Action: Promote high-confidence patterns to SOPs. Archive low-use patterns.
• Keep: Active patterns with recent usage

daily/ logs

• Threshold: 30 days of logs
• Action: Compress months older than current → memory/ARCHIVES/daily-YYYY-MM.md
• Keep: Last 30 days

Memory Sync Protocol (File ↔ Supabase)

When Supabase is available:

Session Start

1. Read local memory files
2. Query Supabase for any updates from other agent interfaces
3. Merge: Supabase wins on newer timestamps
4. Update local files with merged data

Session End

1. Write session data to local files
2. Sync to Supabase memories, session_logs, daily_logs tables
3. Confirm sync success

Conflict Resolution

• Newer timestamp wins
• Higher confidence score wins (for memory facts)
• If truly ambiguous: flag for CC's review

Memory Hygiene Checklist (Monthly)

[ ] SESSION_LOG.md under 200 lines
[ ] ACTIVE_TASKS.md under 50 items
[ ] MISTAKES.md under 30 entries (no duplicates)
[ ] PATTERNS.md under 30 entries (no duplicates)
[ ] LONG_TERM.md — all facts still valid? Confidence scores current?
[ ] SOP_LIBRARY.md — all SOPs still relevant? Success rates accurate?
[ ] daily/ — older than 30 days archived?
[ ] brain/ files — under 2500 lines combined?
[ ] No orphaned references (links to deleted files)?
[ ] Supabase sync current (if available)?

Obsidian Links

• [[skills/INDEX]] | [[brain/CAPABILITIES]]
• [[memory/ACTIVE_TASKS.template]] | [[memory/SESSION_LOG.template]]

---

Maven-specific adaptation

Maven's brain/ houses MARKETING_CANON, attribution model, brand profiles per client, BENCHMARK, OKRs, RISK_REGISTER. memory/ houses ACTIVE_TASKS, SESSION_LOG, MEMORY_INDEX. data/pulse/cmo_pulse.json is rotated continuously. Per-client knowledge is pruned by client lifecycle (active → maintenance → archived). Marketing canon entries are immutable once added; client profiles update freely.