Files
open-memory/AGENTS.md
glm-5.1 f822d93662 Collapse 8 tools to 2: memory router + memory_compact
All read-only operations now dispatch through a single memory tool that
routes by {tool: "name", args: {...}}. This cuts visible tool count from
8 to 2, significantly reducing context bloat for small models. The
memory_compact tool remains separate as it is a write/mutation operation.
2026-04-21 11:17:34 +00:00

7.1 KiB

AGENTS.md

Project

@alkdev/open-memory — an OpenCode plugin that gives agents access to their own session history, context window awareness, and compaction control.

Repository

  • Git: git@git.alk.dev:alkdev/open-memory.git (Gitea, mirrors to GitHub on release)
  • License: Apache-2.0
  • Runtime: Bun
  • Language: TypeScript (strict, ESM, verbatimModuleSyntax)
  • Linter: Biome (bun run lint, bun run format)
  • Build: bun run builddist/ (bun build + tsc declarations)

Commands

bun run build        # bun build src/index.ts + tsc --emitDeclarationOnly
bun run typecheck    # tsc --noEmit
bun run lint         # biome check .
bun run format       # biome format --write .
bun run test         # bun test

Always run bun run typecheck and bun run lint after changes.

Architecture

Three Pillars

  1. Context Awareness — tracks context window usage via SSE events, injects status into system prompts
  2. Session History — read-only queries against OpenCode's SQLite DB using bun:sqlite (readonly mode)
  3. Compaction Management — improved compaction prompt + on-demand compaction triggering

Source Structure

src/
├── index.ts              # Plugin entry: hooks + tool registration
├── tools.ts              # Tool definitions (memory_*)
├── context/
│   ├── tracker.ts        # SSE token tracking (per-session context usage)
│   ├── thresholds.ts     # Threshold constants + ContextStatus type (single source of truth)
│   └── notify.ts         # Context notification formatting
├── history/
│   ├── queries.ts        # bun:sqlite read-only query helper (all DB access goes here)
│   ├── format.ts         # Markdown rendering for session/message output
│   └── search.ts         # LIKE-based full-text search across conversations
└── compaction/
    └── prompt.ts         # Compaction prompt template (self-continuity, not "for another agent")

Plugin Hooks

Hook Purpose
experimental.session.compacting Replace default "summarize for another agent" with self-continuity prompt
experimental.chat.system.transform Inject context % used + advisory into system prompt
event Feed SSE events to ContextTracker

Tools (2)

Tool Purpose
memory Router for all read-only operations: summary, sessions, messages, search, compactions, context, plans, help. Call with {tool: "help"} to see available operations.
memory_compact Trigger compaction via ctx.client.session.summarize() — kept separate because it's a mutation

The memory tool dispatches to internal handlers by tool name, keeping the agent's visible tool count low (2 instead of 8) to minimize context bloat.

Internal operations (accessed via memory({tool: "...", args: {...}})):

Operation Purpose Key args
help Show available operations, or details for a specific one tool (optional)
summary Quick counts: projects, sessions, messages, todos
sessions List recent sessions, optionally filtered by project limit, projectPath
messages Read messages from a specific session sessionId, limit
search Text search across all conversations (LIKE-based) query, limit
compactions List/read compaction checkpoints for a session sessionId, read
context Current context window usage (% , tokens, model, status)
plans List or read saved plan files read (filename)

Database Access

  • Uses bun:sqlite native driver — no subprocess, no sqlite3 CLI dependency
  • Read-only: new Database(path, { readonly: true, create: false })
  • Connection is lazy-initialized and cached (singleton)
  • All queries use db.prepare(sql).all(params) — never string interpolation
  • DB path: ${XDG_DATA_HOME:-$HOME/.local/share}/opencode/opencode.db

Context Tracking

  • Listens to message.updated SSE events for assistant messages
  • tokens.input on the latest assistant message = current context size
  • Compares against model's context limit from config
  • Thresholds: green (<70%), yellow (70-85%), red (85-92%), critical (>92%)
  • System prompt injection at yellow/red thresholds with advisory

Context Percentage Calculation

From OpenCode source (overflow.ts):

count = tokens.total || (input + output + cache.read + cache.write)
reserved = config.compaction?.reserved ?? min(20000, maxOutputTokens)
usable = model.limit.input ? model.limit.input - reserved
                        : model.limit.context - maxOutputTokens

The tokens.input on the last assistant message approximates context size. We track against model context limit from config, falling back to 200k.

Compaction Data in DB

When compaction occurs, OpenCode creates:

  1. A synthetic user message with a compaction-type part (part.data = {type: "compaction", auto: true/false, overflow: true/false})
  2. message.data.summary = {diffs: [...]} on the compaction message
  3. The assistant message immediately after contains the actual summary text in a text-type part

The memory_compactions tool queries for compaction-type parts and retrieves the adjacent summary text, presenting them as navigable checkpoints.

Write Operations

All write operations (compaction triggering) go through the OpenCode client SDK (ctx.client.session.summarize). The plugin never writes to the database or any OpenCode files.

Key Conventions

  • No comments unless requested
  • ESM with .js extension in imports
  • bun:sqlite for all database queries (never spawn sqlite3)
  • Parameterized queries only (never interpolate user input into SQL)
  • Read-only DB access — writes go through the SDK
  • Strict TypeScript with verbatimModuleSyntax
  • Biome for linting and formatting

Relationship to open-coordinator

  • open-coordinator (/workspace/@alkimiadev/open-coordinator): worktree orchestration, session spawning, anomaly detection
  • open-memory: session introspection, context awareness, history browsing
  • Both use SSE events but for different purposes
  • Both implement experimental.session.compacting — open-memory's version is more detailed
  • Can be used together or independently

Implementation Phases

Phase 1: Foundation (current)

  • Plugin scaffolding, build, basic hooks
  • Compaction prompt, context tool, history tools
  • bun:sqlite read-only queries

Phase 2: Context Awareness

  • SSE token tracker (implemented, needs field testing)
  • Proactive system prompt injection
  • memory_compact tool

Phase 3: History Browser Enhancements

  • FTS5 virtual table support (if available)
  • Better search (stemming, ranking)
  • Session comparison tools

Phase 4: Polish

  • Configurable thresholds
  • Export/import helpers
  • Integration tests

References

  • OpenCode source: /workspace/opencodepackages/opencode/src/session/compaction.ts, overflow.ts
  • OpenCode plugin SDK: /workspace/opencode/packages/plugin/src/index.ts
  • Plugin types: see Hooks interface for all available hooks
  • Bun SQLite docs: https://bun.com/docs/runtime/sqlite
  • OpenCode DB schema: message, part, session, project, todo tables