Skip to main content
The core tools every connected agent shares — a remember / recall / list / resume / forget surface, and the entire MCP tool set (pure memory CRUD). They operate on your context memory and are scoped to you on every call.

​
save_context

Save something worth remembering across tools and sessions. Idempotent — saving the same content twice returns the same id with deduped: true.

​
Parameters

​
content
string
required
The thing to remember.
​
memory_type
string
default:"fact"
The canonical memory type, which drives retrieval behavior:
  • fact — atomic, stable knowledge (“the project uses GraphQL”).
  • event — something that happened at a time (a deployment, a decision).
  • instruction — how to do something (a procedure, workflow, runbook).
  • task — what’s being worked on right now. Ephemeral: it auto-expires after a TTL (7 days) and is swept on the next write. It’s still embedded, so it’s semantically retrievable while live.
Defaults to fact (a kind of decision defaults to event).
​
kind
string
default:"note"
Finer-grained display flavor: fact, decision, preference, or note.
​
topic
string
A stable topic this entry is about (e.g. deploy-schedule, editor-preference). A newer entry with the same topic supersedes older ones — latest wins. Use it when updating a fact the user stated before.
​
source_surface
string
Where this came from, e.g. claude-code, chatgpt, cursor.
​
tags
string[]
Optional freeform tags.

​
Returns

​
id
string
The id of the saved (or matched) entry.
​
deduped
boolean
true when the content already existed and was deduplicated.
Example result
{ "id": "ctx_8f2a…", "deduped": false }

​
save_session

Persist what a whole session taught — comprehensively. Pass the transcript or a digest and OneLamp distills it into the reusable facts, events, instructions, and tasks worth keeping, classifies each, and saves them (deduped). This is the comprehensive backstop to per-fact save_context, and matters most for tools without lifecycle hooks. Idle chatter and secrets are skipped.

​
Parameters

​
transcript
string
required
The session transcript, or a concise digest of everything durable the user revealed.
​
source_surface
string
Which tool this came from, e.g. claude-code, codex, cursor.

​
Returns

​
saved
number
New entries saved.
​
deduped
number
Items already known.
​
total
number
Items distilled from the session.
​
session_id
string
The id of the recorded session digest (present when anything durable was saved), for later resume_session.

​
get_context

Retrieve relevant context as a ranked pack of chunks (semantic + keyword + exact fact-key retrieval, fused with weighted RRF). Returns source material to reason over — not an answer. An empty store returns an empty pack, never an error.

​
Parameters

​
query
string
required
What to retrieve context about.
​
max_results
integer
default:"8"
Maximum context chunks to return (1–50).
​
memory_types
string[]
Restrict results to these memory types (fact, event, instruction, task) — e.g. just instruction for runbooks. Omit to retrieve across all types.
​
scope
string
default:"all"
Which context store(s) to read from: personal (your own), team (every team you belong to), team:<id> (one specific team you’re a member of), or all (personal + your teams, fused with personal winning ties). See Teams.
​
verified_only
boolean
default:"false"
If true, return only verified team learnings (ones a 2nd teammate has independently corroborated), dropping unverified and contested ones. Your personal context is always included. Default returns all, trust-weighted.
​
recency
boolean
default:"false"
If true, strongly favor the most recent context over the most relevant — for “what was I just working on” / “what’s the latest” reads. Default is relevance-ranked, with recency only as a tie-breaker.

​
Returns

​
packs
object[]
The ranked context chunks. Each pack has:
Example result
{
  "packs": [
    {
      "id": "ctx_8f2a…",
      "text": "Use pnpm, never npm",
      "kind": "preference",
      "memory_type": "fact",
      "source_surface": "claude-code",
      "ts": 1733500000000,
      "score": 0.8123
    }
  ]
}

​
list_context

Browse your saved memory, newest first — each item’s id, title, kind, memory_type, category, a short text preview, and cross-agent reuse (reuse_count and last_used_surface — how often it’s been retrieved and the tool that last used it). Use it to see what’s stored (e.g. to find an id to forget). For relevance-ranked retrieval use get_context.
​
limit
integer
default:"50"
Maximum items to list (1–200).

​
resume_session

Pick up cross-tool work where you left off. Each session digest is recorded by save_session in any tool, so this is the cross-agent handoff path. Without a session_id, returns your most recent sessions (recency-ranked). With one, loads that session in full — its label/summary and the learnings it produced.
​
session_id
string
Load one session in full by id (from a prior resume_session / save_session).
​
limit
integer
default:"5"
How many recent session digests to list (1–50). Ignored with session_id.
​
since
number
Only list sessions at/after this epoch-ms timestamp.
​
source_surface
string
Only list sessions from this surface, e.g. claude-code, cursor, codex.

​
forget_context

Delete one saved memory by id — completely and immediately (the index row, the R2 source, and the vector). Get the id from get_context or list_context. To instead replace a fact that changed, prefer save_context with a stable topic (supersession keeps history); use forget_context to remove for good.
​
id
string
required
The id of the memory to delete.
​
confirm
boolean
default:"true"
Pass false to preview the target without deleting it — returns what would be removed so you can confirm first.
​
forgotten
boolean
true if an entry was deleted; false if the id wasn’t found (or on a preview).
These tools embody OneLamp’s core principle: writes can dedupe and supersede, reads are retrieval-only, and your data is always exportable (export lives in the OneLamp web app, not as an agent tool). See Privacy & data ownership.