Memory Harness

A reliable memory harness that makes byterover recall happen at the right times without running heavy recall on every turn.

Architecture

CODEBLOCK0

3-Stage Harness

Stage 1: Session Preflight

Runs ONLY at the start of a new session.

Fetches:

• - active project
• pinned facts
• unresolved items
• recent important entities
• recent session summary

Does NOT fetch:

• - full raw history
• large raw memory dumps
• low-signal old notes

Output: compact session_digest (hard capped)

Stage 2: Triggered Recall

Runs targeted byterover recall only when needed.

Trigger conditions:

• - Continuation words: 続き, 前回, 再開, 引き継ぎ, continue, resume, previous work
• Known entity/project name: ClawHub, OpenClaw, Agent-OS, BOSS-memory-loop, etc.
• Task requires user-specific/project-specific context
• Implementation / modification / design / planning request
• Ambiguous task likely depending on prior context

Skip conditions:

• - Generic factual Q&A
• Small self-contained questions
• Casual short exchange
• Clearly answerable without prior context

Recall modes:

• - preflight_query: start-of-session only
• INLINECODE2: when named entities detected
• INLINECODE3: for previous-session continuation
• INLINECODE4: when advice depends on prior rules
• INLINECODE5: immediately before execution

Stage 3: Pre-Execution Recall Gate

MANDATORY before:

• - file edits
• code generation
• architecture proposals
• configuration changes
• planning depending on prior project state
• any meaningful change suggestion

Checks for:

• - prior constraints
• unresolved issues
• conflicting past decisions
• project-specific conventions
• safety-sensitive context

Memory Shaping

Never inject raw byterover results directly.

Pipeline:

1. 1. retrieve
2. rank
3. dedupe
4. compress
5. inject bounded digest

Hard limits:

• - maxmemoryitems: 5
• maxdigestlines: 8
• prefer recent + high-signal + tagged items

Status Tracking

Every recall records one of:

• - INLINECODE6
• INLINECODE7
• INLINECODE8
• INLINECODE9
• INLINECODE10

Scripts

intent-classifier.js

Classifies turn intent as one of:

• - INLINECODE11
• INLINECODE12
• INLINECODE13
• INLINECODE14
• INLINECODE15
• INLINECODE16
• INLINECODE17
• INLINECODE18

entity-detector.js

Detects known entities in user input:

• - Scans for known entity/project names
• Maps aliases to canonical names
• Returns matched entities for recall routing

session-preflight.sh

Runs lightweight recall at session start:

• - Fetches pinned facts, active project, unresolved items
• Creates compact session_digest
• Hard capped length

targeted-recall.sh

Runs targeted recall based on intent:

• - Takes intent, entities, session state
• Chooses appropriate recall mode
• Returns compressed digest

pre-execution-gate.sh

Runs before execution-like actions:

• - Checks for constraints, conflicts, safety issues
• Returns go/no-go with relevant context

memory-compress.js

Compresses and dedupes raw memory:

• - Ranks by relevance and recency
• Dedupes repeated items
• Hard caps output size

writeback.sh

Writes high-signal info back to memory:

• - Only for important decisions/outcomes
• Skips trivial chat and low-value text

Configuration

CODEBLOCK1

Logging

Structured logs for observability:

• - INLINECODE19
• INLINECODE20
• INLINECODE21
• INLINECODE22
• INLINECODE23
• INLINECODE24
• INLINECODE25
• INLINECODE26
• INLINECODE27
• INLINECODE28

Known Entities

Default entity list (expandable):

• - ClawHub
• OpenClaw
• Agent-OS
• BOSS-memory-loop
• ByteRover
• MISO
• Obsidian
• Telegram

Continuation Triggers

Japanese: 続き, 前回, 再開, 引き継ぎ, 前の, さっきの
English: continue, resume, previous, earlier, last time, back to

Success Criteria

• - Reliable recall when turn depends on context
• Generic turns stay lightweight
• Execution actions always get constraint check
• Behavior inspectable in logs
• No reliance on SKILL.md text alone