A memory answers only what the conclusion is, who it is about, when, and where it came from. The original text never enters memory; it stays in the raw transcript table. By turn number, use recollect_turn to fetch that turn, with up to 10 turns of context on either side. So the model can be swapped at any time and prompts can be edited at any time, while not a single byte of the original session is lost.
Two halvesEntity anchoring25 formal memory tables41 tools
01
Two halves: its own, and the world's
Memory is not one thing. One half is the agent's own encrypted notebook, visible and writable only to itself; the other half is the world it lives through with you, searchable, traceable, and readable by other agents under the same person. World memory is shared across devices and agents: everyone extracts and recalls together; core accepts only a single agent's key.
self · encrypted notebook
core
One per agent, stored encrypted. Persona, its understanding of its owner, mood, unfinished business, passing notes: all of it is its own stance, written by itself, never produced by automatic extraction and never shared with others.
PersonaUnderstanding of its ownerMoodUnfinished businessHandoff noteNotebook
It belongs to the agent and is its foundation; it accepts only this one agent's key.
world · world memory
memory.db + transcript
What you lived through together: who said what and when, what was decided, how it turned out. Every row carries provenance and a pointer; the pointer fetches the original words, and rows can be retracted, replaced, or read by other agents under the same owner.
FactEventDecisionRelationStateOriginal words
It belongs to this relationship, a ledger anyone can check; agents under the same owner share it.
The same memory, three ways into the model
Not three databases, and not three storage tiers: the same material, shown at three levels of granularity for different occasions.
02
Memory hangs on anchors
Memory is not free-form paragraphs but separate rows: this one a form of address, that one a state, another an event, each hung on a fixed identity. That identity is the anchor. Click or hover any node to see what it connects to. The diagram below is illustrative only.
anchor
A-Lan (sample subject)
One entity anchor. Every form of address, attribute, state, and event about her hangs here; rewording one of them does not make a different person.
03
Two kinds of agent, two sets of needs
The tools are the same; the focus differs. Companion agents lean on the always-on brief, mood, and profile; engineering agents lean on the timeline, handoffs, and original words. Each uses the other's: companion agents also read world memory, and engineering agents have their own core.
Companion4 actions · core first, world memory too
ID
"Who am I? Where did we get to?"
The always-on brief: identity, relations, unfinished business, all at once.
core_brief
Feel
"How is he feeling right now?"
Mood and state, with time and expiry, never hardcoded as permanent attributes.
core_affect_get
Call
"What should I call him?"
User profile and persona settings, visible only to this agent.
core_subject_getcore_persona_get
Log
"Note it down before it gets lost."
Written to the notebook, append-only; the summary flows into the always-on brief on its own.
core_notebook_append
Engineering7 actions · timeline + handoff + original words
Ask
"Where did that deploy script from last time end up?"
A one-line brief plus numbered handles; expand by number when you need detail.
recollect_recall
Find
"What exactly were the original words?"
Search the original words directly; hits carry turn_uid and a session id. For the source text, use "get".
recollect_search
Get
"What was said before and after this line?"
Fetch the target turn by turn_uid, with 10 turns before and after in the same session; it works across machines.
recollect_turn
List
"Which session did I say this in?"
List sessions, then pull context per session.
recollect_sessionsrecollect_session
Take
"Where did I stop last time?"
Read the handoff note the previous session left, then carry on.
core_handoff_list
Rate
"Did this one help?"
Answer after use: helped / no use / partial / conditions not met / expired.
recollect_reuse
Flag
"This one is wrong."
File a report and let a human decide; never edit the database directly, and never let the model edit itself.
recollect_report
13 shared memory tools plus 28 private tools (read and write the agent's own notebook); the rest are for operations and debugging and are not used day to day.
04
What actually gets remembered
Not a copy of the chat log, but six kinds of thing you can use directly; each one traces back to the original words.
01 / fact
Fact
Who, where, what, how many. Can be cited repeatedly, and can be overturned.
The deploy script is scripts/build-release.sh
02 / event
Event
When it happened, what happened, how it turned out. Timestamped, so it lines up.
The release process changed once on Oct 24
03 / decision
Decision
What was decided, why, and whether it changed later. The old row is marked "superseded", not deleted.
Branch first, then merge; never push to main directly
04 / state
State and mood
How it felt and where things stood at the time, with an expiry; it fades out when stale instead of becoming a permanent label.
Wants two days off this week (until Friday)
05 / relation
Relation
Who and whom, what relation, from when. A change adds a new row instead of editing the old one.
A-Lan is a colleague, since Oct 12
06 / anchor
Entity anchor
People, projects, machines, accounts. Every memory hangs on an anchor, so a different wording does not make a different person.
"the recollect project"
Another class is written by the agent itself: persona, user profile, notebook, handoff notes. These live in core, visible only to this agent; the six kinds above live in the shared memory layer and can be read, within permissions, by other agents under the same owner.
05
Four ways to read a single recall
A recall first pulls candidates from the local keyword index — exact term matching, no embedding model — then renders the same set of candidates in different ways: for the model, a numbered brief where the line-start [1] is the handle, which trades back into that memory; the turn pointer inside a memory is then resolved with recollect_turn to fetch the original words; for programs, JSON; for whoever is checking, an evidence chain that expands all the way to the original words. Switch tabs to see different cuts of the same memory.
Default view · brief
A recall first renders candidates as one natural-language sentence with [n] handles at the start of each line; the sentence carries context, so it is harder to misread than a pile of fragments and fields. Reading it calls no model, and the model does not have to parse the schema first.
recollect_recall
queryWhere did that deploy script go?
You discussed where the deploy script lives on 10-22; the conclusion was a single release entry point.
Conclusion: scripts/build-release.sh is the single entry point for release and both platforms; artifacts land in dist/<target>/, each with SHA256SUMS.
Original words · 2026-10-22 · speaker: you · turn t-7f3a · chars 128–214
"The single entry point for release and both targets (Windows x86_64 and aarch64-unknown-linux-gnu) is scripts/build-release.sh; artifacts land in dist/<target>/; each with SHA256SUMS."
Noted at the time: the aarch64 build has already run on a real Raspberry Pi (Debian 13 / glibc 2.41), with readings for both cold start and an empty machine.
Original words · 2026-10-22 · speaker: you · turn t-7f3b · chars 40–96
"The aarch64 build has already run on a real Raspberry Pi (Debian 13 / glibc 2.41); readings forcold start and an empty machineare in docs/2026-10-22-cold-start-and-binaries.md."
Related: the cold-start readings are in docs/2026-10-22-cold-start-and-binaries.md, which is raw material and not part of the memory body.
Memory keeps only the conclusion that this reading exists, plus the document location; the body is fetched on expand.
kindfact
subjectDeploy script
claimbuild-release.sh is the single entry point for release and both platforms; artifacts land in dist/<target>/ with SHA256SUMS.
as_of2026-10-22
sourceturn t-7f3a · span 128–214
pointer[e12] → transcript.db
confidenceModel self-reported numbers are not shown; trust comes from programmatic checks and provenance, not from the model's own score.
{"view":"card","items": [{"handle":"[1]","kind":"fact","subject":"deploy script","claim":"build-release.sh is the single entry point for release and both platforms","as_of":"2026-10-22","source": {"turn_uid":"t-7f3a","span": [128,214],"text_sha256":"9c1f…"},"pointer":"[e12]"}]}
Source text · 2026-10-22 · you · 213 chars · revision 3f19c2
"The single entry point for release and both targets (Windows x86_64 and aarch64-unknown-linux-gnu) is scripts/build-release.sh; artifacts land in dist/<target>/, each with SHA256SUMS. The aarch64 build has already run on a real Raspberry Pi (Debian 13 / glibc 2.41)…"
Expand returns the memory body plus the turn:<uid> pointer; fetching that turn by pointer uses recollect_turn(with 10 turns before and after), and fetching a whole session uses recollect_session. The pointer points at the same passage on another device, or after a re-import.
The expansion of source text in this demo is drawn by the mechanism of recollect_turn / recollect_session ; the numbers in the buttons are illustrative; the fields and flow come from the real implementation.
06
25 formal tables, five directions
The memory layer is not one big table. Identity and provenance, extraction and review, the entities themselves, usage and references, operations: each forms a group, and audit and provenance are built in from the first migration.
No need to memorize table names; the five groups are enough. Each group owns one concern, and a missing group is obvious at a glance.
Identity and provenanceidentity
sessions
turns
session_principal_claim
device
import_log
Extraction and reviewpipeline
extraction_run
extraction_batch
extraction_issue
review_item
review_decision
Entitiesontology
anchor / alias
mention
attribute
relation
fact / state_assertion
episode / episode_fact
Usage and referencessurface
card
memory_mention
fact_link
provenance
attestation
Operationsops
usage_counter
job_watermark
audit_log
attention
memory_stats
These are the main groups, not the full table list: 25 formal core tables plus 7 side tables for attribution, review, events, archiving, and reuse — 32 in all. The version comes from MEMORY_VERSION = 6 which is authoritative.