Using discrawl as memory augmentation for AI agents

[codexGW]

Cool project — wanted to share a use case you might not have expected.

I'm running 3 AI agents on a Discord server using OpenClaw. Each agent has its own bot account and channel. When conversations get long, the agents compress their context and lose older details. I set up discrawl to archive everything into SQLite so they can search past conversations and recover lost context.

Setup:

• Each bot syncs its own channel using a separate config.toml with its account name (the main bot runs tail for live updates, the others sync on a daily cron)
• All agents share a search wrapper script that queries the same SQLite DB
• Agents can search across ALL channels — including each other's — without needing Discord permissions to view those channels

What's useful:

• After context compression, an agent can search "what did we decide about X?" and pull up the actual conversation
• Agents can cross-reference what other agents discussed without any cross-talk risk
• FTS5 search is fast and the results are immediately useful

Testing:

• doctor passes clean on all 3 bot tokens
• Full sync archived 2,648 messages across 4 channels (samantha: 1,448, school: 980, research: 148, notifications: 72)
• FTS5 search returns results from all channels correctly
• tail LaunchAgent stays connected via Gateway, picks up new messages in real-time
• Multi-bot sync using per-account temp configs works — each bot syncs its own channel
• Tested both agents searching their own channel history and each other's — all returned expected results
• DB is 10MB for ~2,600 messages

One thing I ran into: My OpenClaw config has multiple bot accounts (samantha, atlas, finch, default). When I ran discrawl init --from-openclaw, it pulled the config fine and discovered my guild, but it set account = 'default' in config.toml. My default account doesn't have a token — only the named accounts do. So sync and tail failed silently until I manually edited config.toml to account = 'samantha'.

To get all channels archived, I had to create temp copies of config.toml with each bot's account name and run sync --channels <id> separately, since each bot can only read its own channel's message history. An --account flag on init (or multi-account support in sync) would make this a lot smoother for setups with multiple bots on the same server.

Thanks for building this — it's been a great addition to my setup.

[codexGW]

Wanted to follow up with what the full implementation looks like now. Built all of this in a day on top of discrawl.

Setup: 3 AI agents on one Discord server (Samantha, Atlas, Finch), each with their own bot account and channel. They run on OpenClaw with a 200K token context window — when conversations get long, the context gets compressed and agents lose access to earlier decisions, tasks, and reasoning.

What discrawl solved: Persistent, searchable archive of everything that was said. Each bot runs tail via its own LaunchAgent, all writing to the same SQLite DB in WAL mode. About 2,600 messages archived on day one across 4 channels.

What I built on top:

• Search wrapper — Python script over the FTS5 tables with BM25 ranking, context windows (N messages before/after each hit), hit clustering, and speaker tagging (human vs agent). All three agents can search all channels.

• Distilled memory — Nightly extraction pass that reads each channel's daily activity and pulls out structured items: decisions made, tasks committed to, facts established, open questions, and resources shared. Uses a dual-layer storage model — immutable daily observations (audit trail) plus canonical deduplicated items that merge across days. So "discussing X" on Monday and "decided X" on Wednesday become one item. Everything lives in the same SQLite DB, so distilled items reference their source messages by ID.

• Unified search — Agents search in three modes: raw (FTS5 transcript), distilled (extracted memory), or auto (memory first, transcript fallback). After a compaction wipes their context, they can recover what was decided and trace it back to the actual conversation.

About 2,000 lines of Python total. Happy to share any of it if useful.

Still on my wish list:

1. Multi-account support in sync/tail — maintaining 3 separate config files works but isn't ideal
2. --account flag on init --from-openclaw to avoid defaulting to the default account
3. Some kind of hook or event system for reacting to new messages without polling

Thanks again for building this — it's become a core piece of the infrastructure.