Memory

Your window into everything an agent knows or should know. Sessions, daily notes, soul / rules / skills, dreams, checkpoints. All indexed, hybrid scored, ranked. You and your agents both reach for it when you need to know what was decided, where a rule came from, or why an agent is acting weird today.

The runtime owns the data, this just shows you what’s there.

The memory dashboard: tier overview cards on top, search and filters in the middle, a recent feed (or your search hits) below.

Search

Type a query, get ranked hits across every tier in one list. Hybrid scoring (keyword + semantic) so “launch plan” finds the document titled exactly that AND the session where you talked about going to market. Click any row for the full content in a side drawer.

Search across every tier in one ranked list, with the agent strip and tier chips above the results.

Three filters narrow the result set:

• Agent: avatar strip on the left. Click an avatar to scope to one agent.
• Tier: multi-select chips. Pick sessions + daily_note to focus on conversational history, durable for the canonical layer, etc.
• Kind (under durable only): slice the durable layer further into soul, rules, skills, tools, identity, memory-log, dreams, user, bootstrap.

What it’s good at:

• “What did we decide about the brand voice?” → durable + sessions + daily notes, scoped to the agent that owns it.
• “Why is Roscoe acting weird today?” → clear the search, agent filter set to Roscoe, scan the recent feed.
• “Where did this rule come from?” → tier durable, kind rules, search the keyword.

When the search bar is empty, the page shows a recent feed across the same filters so you can browse without a query. Overview cards up top give live row counts per tier.

What’s in there

Tier	What it holds
Sessions	Chat threads, by agent. Token counts, model, status, and the back-link to the conversation itself.
Daily Notes	End-of-day summaries an agent writes for itself. Quick way to scan what’s been on their mind without scrolling through every session.
Durable	Long-term knowledge. Soul, identity, rules, tools, skills, MEMORY-LOG, the stuff that defines who the agent is. Filter by Kind to zoom in on just the soul, just the skills, just the rules.
Dreams	Background reflection an agent does between active work. Phase docs, signals, the agent’s own theories about what’s going on.
Checkpoints	Compaction snapshots. When a session gets too big, the agent compresses it into a checkpoint and starts fresh. The checkpoint is what carries forward.
Turns (System Logs)	Every individual message inside a session.
Audit (System Logs)	Every Bakin event in the system.

Turns and Audit are noisy by default. Flip the System Logs toggle in the page header to bring them into the dashboard.

Settings

Setting	Type	Default	What it does
Backfill window (days)	number	30	On first activation, index this many days of history across all tiers.
Skip sessions over (bytes)	number	10 * 1024 * 1024	Transcripts larger than this are skipped to keep the indexer responsive.
Skip .reset backup transcripts	boolean	true	Historical reset backups are not live state; skip by default.
Compare against runtime recall	boolean	true	Show runtime daily-note recall alongside Bakin search results.
Turn retention (days)	number	7	Turns older than this are dropped at write time and pruned daily. The runtime still owns the source transcript.
Audit retention (days)	number	30	Audit rows older than this are dropped at write time and pruned daily.

Defaults are sensible. Most folks will never touch this.

HTTP API surface for this plugin: see the API reference.

For agents

Agents query memory through MCP exec tools.

• bakin_exec_memory_get_session: Fetch a session by key plus its most recent turns.
• bakin_exec_memory_get_turn: Fetch a single turn by id (the turn:<hex> form).
• bakin_exec_memory_list_agents: Agents with memory rows, each with total count and per-tier breakdown.
• bakin_exec_memory_search: Hybrid search across every memory tier (sessions, turns, checkpoints, daily notes, dreams, durable, audit). Optional tier/agent filters.
• bakin_exec_memory_status: Indexer health: per-tier row counts, offset tracking, snapshot timestamp.

Full schemas in the Exec tools reference.

Related

• Essentials: cross-table search includes memory automatically
• Team: the agents whose memory tiers Memory indexes
• Activity Feed: live event stream that audit memory replays after the fact