Dory

Minimal agent with forgetting (500 lines).

Agents today are bad at forgetting, and worse at remembering that they forgot. Forgetting is a good thing actually. Evolution didn't select for photographic memory in humans for a reason: when you're planning a trip to Japan and you're picking a seat on the flight, you don't need sushi restaurants in your head — and later, when you're planning meals, you can rehydrate that context on demand. Dory does the same: old context gets compacted — details fade but the gist remains, and the original is always one UNCOMPACT away.

agent.py             — the loop
SOUL.md              — purpose + tool format (immutable)
SEGMENT.md           — chunking instructions for compactor (agent-editable)
SUMMARISE.md         — summarisation instructions for compactor (agent-editable)
blobs/               — content-addressed blob store (shared)
chats/<id>.md        — chat rooms (one file per room, append-only)
agents/<self>/       — per-identity dir; <self> = hash(SOUL + boot ts)
  ├ LIFE.md          — append-only event log (every turn, every tool)
  ├ MEMORY.md        — agent-editable working notes (capped at 4000ch)
  ├ context.md       — last live context (resume point)
  └ subs/            — chat subscriptions (symlinks into chats/)

Design

• SOUL / LIFE / MEMORY. SOUL is the agent's purpose (immutable). LIFE is the append-only event log, per-identity; the last 50 lines are injected each turn. MEMORY is the agent's own scratchpad — learned preferences and notes it edits with EDIT_FILE.

• SEGMENT / SUMMARISE. Instructions for the compactor, stashed as blob refs in the agent's prompt (so they don't bloat it). The agent can UNCOMPACT or READ_FILE them, and even edit them to tune its own compaction strategy.

• Prompt order. Each turn: `SOUL + segment_ref + summarise_ref

  • harness + context + MEMORY + LIFE`. Stable prefix first, append-only context next, volatile tail last — optimised for LLM prompt caching.

• Blobs are the substrate. Every LLM response and every tool result is stashed as a blob under blobs/; the ref is appended to LIFE.md. Blobs are content-addressed (filename = sha256 prefix of the blob) and have a tiny frontmatter: at, gist, parent. Live memory holds content inline for free access; older content may only appear as a ◱hash=... gist=...◲ reference after compaction.

• Chunked compaction. When context exceeds the limit, compaction runs in the background. Phase 1: an LLM call identifies conceptual chunks in the first 2/3 of context. Phase 2: each chunk is compacted in parallel — the compactor sees the full agent state (soul, harness, memory, life) and returns a summary, a relevance score (1–10), and a gist. Chunks are swapped in from least to most relevant until eligible content is compressed to 50%. The recent 1/3 of context is never touched. Originals are stashed as blobs with LLM-generated gists; UNCOMPACT rehydrates any ref inline.

• Graceful forgetting. Compaction is semantic, not truncation. Old turns lose resolution but stay reachable via blob refs. The agent can UNCOMPACT any ref to pull it back into context.

• Chat = file. Each room is chats/<id>.md, append-only. Lines are tagged [USER:name ts] or [AGENT:name ts]. Agents subscribe by symlinking files into agents/<name>/; the harness tails that dir each turn. SAY <id> <msg> appends to the right file. External transports (Telegram, stdin, etc.) are independent bridge processes that read/write the same files — the agent doesn't know or care.

Running

pip install -r requirements.txt
export DEEPSEEK_API_KEY=...
python agent.py            # fresh boot: new identity
python agent.py <hash>     # resume an existing agent by its <self> hash

Defaults to deepseek-v4-pro via api.deepseek.com. Reasoning content is captured and prepended to context inside <reasoning> tags. Pass --verbose to print LIFE events to stdout.

Each fresh boot generates a new content-addressed identity: the agent stashes a blob of its SOUL plus boot timestamp, and uses that hash as its name. Two agents with the same SOUL booted at different times get distinct identities; their state lives under agents/<self>/.

Edit SOUL.md to give the agent a purpose before running. Talk to a running agent with python chat.py [room_id] (defaults to room dev). Wipe runtime state with python clean.py.