Ralph Chang ff4639d153 fix: PR-2 memory plugin behavior improvements
## Task 5: Canonical exact dedupe
- Already implemented in PR-1 with enforceLongTermLimits()
- Source priority: explicit > manual > compaction
- Same source: higher confidence wins

## Task 6: Structured negative guard
- Add isNegatedMemoryRequest() for adjacency detection
- "不要記住" / "don't remember" are now properly ignored
- "not forget to remember" no longer false positive (not a directive)
- Restrict patterns to line-start (^|\n) to avoid mid-sentence matches

## Task 7: Compaction quality gate
- Add shouldAcceptWorkspaceMemoryCandidate() predicate
- Reject low-quality candidates: git hashes, errors, stack traces
- Reject temporary progress, code signatures, path-heavy facts
- Only accept entries with >= 20 chars

## Pattern improvements
- All patterns use matchAll() with proper g flag
- Dedupe by canonical text in extractExplicitMemories()
- Line-start anchor prevents "to remember" mid-sentence matches
- Add more trigger patterns: save/add to memory, commit to memory

Tests: 36 passing
2026-04-26 13:06:36 +08:00
2026-04-25 18:20:42 +08:00
2026-04-26 12:52:21 +08:00

OpenCode Working Memory Plugin

npm version License: MIT

Advanced four-tier memory architecture that keeps your AI agent sharp, focused, and never forgets what matters.

Stop losing context across compactions. Stop watching your agent repeat the same mistakes. This plugin gives your OpenCode agent a professional-grade memory system that scales with your project complexity.

What You Get

  • 🧠 Core Memory - Persistent goal/progress/context blocks that survive compaction
  • 💡 Working Memory - Smart slot-based system (errors, decisions, todos, dependencies)
  • 🎯 Memory Pressure Monitoring - Real-time token tracking with automatic interventions
  • 🧹 Auto Storage Governance - Prevents unbounded disk growth (300 files/session, 7-day TTL)
  • 📊 Smart Pruning - Pressure-aware tool output compression
  • Zero Configuration - Works out of the box, highly customizable

Installation

Option A: Let an AI agent do it

Paste this into any AI agent:

Install opencode-working-memory plugin by following:
https://raw.githubusercontent.com/sdwolf4103/opencode-working-memory/main/docs/installation.md

Option B: Manual setup

Add to your ~/.config/opencode/opencode.json:

{
  "plugin": ["opencode-working-memory"]
}

Restart OpenCode. The plugin is downloaded and installed automatically.

For LLM Agents

  1. Edit ~/.config/opencode/opencode.json
  2. Add "opencode-working-memory" to the plugin array
  3. Verify the config file was saved correctly

Quick Start

The plugin works automatically once installed. No configuration needed.

Try telling your agent:

Use core_memory_update to set my current goal
Use core_memory_read to show me what you remember
Use working_memory_add to remember this file path

Features

🧠 Core Memory

Persistent blocks that survive conversation resets:

  • goal (1000 chars) - Current task/objective
  • progress (2000 chars) - What's done, in-progress, next steps
  • context (1500 chars) - Key file paths, conventions, patterns

💡 Working Memory

Auto-extracts and ranks important information:

  • Slots (guaranteed visibility): errors, decisions, todos, dependencies
  • Pool (ranked by relevance): file paths, recent activity
  • Exponential decay keeps memory fresh
  • FIFO limits prevent bloat

🎯 Memory Pressure Monitoring

Real-time token tracking from session database:

  • Monitors context window usage (75% moderate → 90% high)
  • Proactive intervention messages when pressure is high
  • Pressure-aware smart pruning (adapts compression based on pressure)

🧹 Storage Governance

Prevents unbounded disk growth:

  • Auto-cleanup on session deletion (all artifacts removed)
  • Active cache management (max 300 files/session, 7-day TTL)
  • Silent background operation

📊 Smart Pruning

Intelligent tool output compression:

  • Per-tool strategies (keep-all, keep-ends, keep-last, discard)
  • Pressure-aware limits (2k/5k/10k lines based on memory pressure)
  • Preserves important context while reducing noise

Documentation

Tools Provided

The plugin exposes these tools to your OpenCode agent:

  • core_memory_update - Update goal/progress/context blocks
  • core_memory_read - Read current memory state
  • working_memory_add - Manually add important items
  • working_memory_clear - Clear all working memory
  • working_memory_clear_slot - Clear specific slot (errors/decisions)
  • working_memory_remove - Remove specific item by content

How It Works

┌───────────────────────────────────────────────────────────┐
│  Core Memory (Always Visible)                             │
│  ┌─────────┬──────────┬──────────┐                        │
│  │  Goal   │ Progress │ Context  │                        │
│  └─────────┴──────────┴──────────┘                        │
└───────────────────────────────────────────────────────────┘
                            ↓
┌───────────────────────────────────────────────────────────┐
│  Working Memory (Auto-Extracted)                          │
│  ┌──────────────────┬──────────────────┐                  │
│  │  Slots (FIFO)    │  Pool (Ranked)   │                  │
│  │  • errors        │  • file-paths    │                  │
│  │  • decisions     │  • recent        │                  │
│  │  • todos         │  • mentions      │                  │
│  │  • dependencies  │  • decay score   │                  │
│  └──────────────────┴──────────────────┘                  │
└───────────────────────────────────────────────────────────┘
                            ↓
┌───────────────────────────────────────────────────────────┐
│  Memory Pressure Monitor                                  │
│  • Tracks tokens from session DB                          │
│  • Warns at 75% (moderate) / 90% (high)                   │
│  • Sends proactive interventions                          │
│  • Adjusts pruning aggressiveness                         │
└───────────────────────────────────────────────────────────┘
                            ↓
┌───────────────────────────────────────────────────────────┐
│  Storage Governance                                       │
│  • Session deletion → cleanup all artifacts               │
│  • Every 20 calls → sweep old cache (300 max, 7d TTL)     │
│  • Silent background operation                            │
└───────────────────────────────────────────────────────────┘

Why This Plugin?

Without this plugin:

  • 🔴 Agent forgets context after compaction
  • 🔴 Repeats resolved errors
  • 🔴 Loses track of project structure
  • 🔴 Context window fills up uncontrollably
  • 🔴 Disk space grows unbounded

With this plugin:

  • Persistent memory across compactions
  • Smart auto-extraction of important info
  • Real-time pressure monitoring with interventions
  • Automatic storage cleanup
  • Pressure-aware compression
  • Zero configuration, works immediately

Does working-memory system increase token usage? It depends.

It depends on your workflow.

  • 🧹 Clean Slate user (for example, using DCP and frequently restarting sessions)

    • ⚠️ Yes, it might add slight overhead.
    • Because you keep starting fresh, automated memory persistence does not get enough time to pay off.
  • 🚀 Long Haul user (staying in one session until token limits/compaction hit)

    • This plugin is a token saver.
    • Without it, compaction can cause the agent to lose the goal, forget active files, or make wrong assumptions, which creates correction loops.
    • By preserving high-value context (Goals, Progress, Active Files), the agent inherits its previous state quickly. The small memory-prompt cost avoids the larger cost of the agent getting lost.

Configuration (Optional)

The plugin works great with zero configuration. To customize behavior, modify the constants at the top of index.ts. See the Configuration Guide for all tunable options.

Requirements

  • OpenCode >= 1.0.0
  • Node.js >= 18.0.0
  • @opencode-ai/plugin >= 1.2.0

License

MIT License - see LICENSE file for details.

Support

Credits

Inspired by the needs of real-world OpenCode usage and built to solve actual pain points in AI-assisted development.

This project is not affiliated with or endorsed by the OpenCode team.


Made with ❤️ for the OpenCode community

Languages
TypeScript 98.1%
JavaScript 1.9%