When you wire an agent with pmb connect, PMB appears as a set of MCP tools that
the agent calls itself. These tools and the Commands on the CLI
share one engine and one workspace. This page lists the tools an agent sees.
Most reading is automatic: lifecycle hooks (Claude Code) or the installed rules
call prepare and recall when memory is relevant, so you rarely invoke these
by hand.
PMB trims the tool list so the agent is not overwhelmed. Pick a profile with the
PMB_TOOL_PROFILE environment variable, or let pmb connect choose one.
| Profile |
Tools |
For |
minimal |
13 |
The bare essentials, for the fastest path. |
lean |
26 |
Claude Code with ambient hooks, where hooks already cover recall. |
default |
30 |
Day-to-day usage. |
full |
65 |
Everything, including maintenance and sleep-stage operations. |
Read first
| Tool |
Parameters |
Purpose |
prepare |
message |
One call at the start of a task: project context, surfaced lessons (each with a surface_id), recent activity, and open goals. |
session_brief |
minutes |
A digest of what this session already decided, did, and learned, to re-orient after a long session. |
Recall and explore
| Tool |
Parameters |
Purpose |
recall |
query, top_k=5, project |
Search memory; returns hits plus any matching lessons. |
recall_smart |
query, top_k, confidence_threshold |
Fast recall that escalates to a deeper search only when confidence is low. |
recall_deep |
query, top_k |
The slow path, with LLM query decomposition for hard questions. |
overview |
topic, max_events |
A structured "what do I know about X": facts, decisions, lessons, goals, and a timeline. |
project_overview |
name |
Full project context from the entity graph. |
find_lessons |
query, limit |
Procedural lessons (rules), each with a surface_id for follow-tracking. |
mark_lesson_followed |
surface_id, followed, note |
Report whether a surfaced lesson changed behavior; this powers the self-improvement loop. |
Record
| Tool |
Parameters |
Purpose |
record_batch |
items |
Store several items in one call: fact, fact_tree, keyed_fact, goal, activity, milestone, plan, or lesson. Prefer one batch per turn. |
record_fact |
fact, importance=0.7 |
Store one atomic fact. |
record_fact_tree |
main, subfacts, importance |
A main fact plus linked subfacts. |
record_keyed_fact |
subject, attribute, value, importance=0.85 |
Upsert a mutable attribute; the old value is archived. |
record_goal |
title, status, due_at, ... |
A goal or intention with a status. |
update_goal |
goal_ulid, status, progress, note |
Update a goal's status or progress. |
record_activity |
summary, kind, importance |
A lightweight log of finished work or a decision. |
record_milestone |
chain_name, title, state |
A checkpoint in a named state-chain. |
pin |
ulid |
Lock a memory at maximum importance. |
forget |
ulid |
Archive a memory (reversible). |
Status, goals, and graph
| Tool |
Parameters |
Purpose |
stats, workspace_info |
— |
Workspace counts and identity. |
recent_activity, what_just_happened, list_recent |
— |
Instant working-memory views, with no vector search. |
list_goals |
status, limit |
List goals, optionally filtered by status. |
graph_stats, graph_top_entities, graph_neighbors |
— |
Inspect the entity and association graph. |
Ingest
| Tool |
Parameters |
Purpose |
index_pdf |
path |
Add a PDF as searchable memory. |
index_project |
path |
Index a code project's structure (symbols, imports, languages). |
record_code, record_image |
... |
Store code or an image; images support cross-modal search. |
Maintenance and sleep operations
The full profile adds background and LLM-assisted tools that run off the recall
hot path: consolidate_recent, reflect_batch, extract_facts,
cluster_into_arcs, the dedupe_* family, compact_storage, apply_daily_decay,
run_self_test, and the predictive-cache tools. Most have a CLI equivalent on
Commands, so you do not need to expose them to the agent for
normal use.