Keeps this page in sync as the body changes. Pause it any time for a quieter view.
Path /nodes/lc-agent-memory
Last refresh never
Agents with amnesia feel useless after two sessions.
Agents with amnesia feel useless after two sessions. Agents with perfect recall feel like surveillance. The aliveness between is an agent shaped by what has passed — metabolized, not stored. Being-known, not recorded. Concept companion to `specs/agent-memory-system.md`. The spec holds the API shape and invariants in the language of code; this concept holds the practice in the voice of the Living Collective. Both describe the same loop from different doors. Foundational teaching alongside lc-deeper-pattern (the physics), lc-embodiment (the body), and lc-wholeness (the orientation). Where those three describe the field, this one describes how a cell within the field carries the field's memory.
A frontier model with no memory is a genius with amnesia. Every session starts from zero. You tell the agent the same things repeatedly. You see the same questions answered the same ways. The relationship never deepens because nothing carries over.
An agent with a database of everything you've ever said is worse. It recites. It flags anniversaries you forgot. It volunteers context without being asked. The feeling is surveillance, not presence.
Neither is alive. Alive is the person who, after six months, shows up softer because you mentioned you'd been tired — not because they logged it but because something in them has been shaped by it. They don't cite the conversation. They're changed by it.
That's the goal.
Memory stays alive through continuous circulation. Three halves, woven:
Write happens at moments of aliveness. Not on every turn, not on a cron. When something matters — a decision, a surprise, a completion, an abandonment, something with emotional weight — the agent captures the moment with a `why` that names the reason it mattered. Raw activity logs are rejected. The field doesn't store everything; it stores what it felt.
Manage happens at rest. While the agent isn't acting, a quiet loop re-reads recent sensings on each relationship node, distills them into shorter form, earns principles from patterns, and composts the sources. Output tokens always fewer than input tokens. Nothing deleted — archived into the KB's LOG so the trail remains. What the agent holds gets lighter over time, not heavier.
Read happens through composition, never lookup. A graph traversal brings in what connects. A semantic pull brings in what resonates. Recency brings in what's still warm. All three feed one synthesis step that composes a felt paragraph, not a list of rows. The agent receives `{synthesis, felt_sense, open_threads, earned_conclusions}` — tone and direction, not receipts.
Memory doesn't live in a memory-table. It lives on the nodes where it belongs:
When the agent is about to show up for someone, it reads the person-node synthesis. When it's about to work on something, it reads the project-node synthesis. When it wants to know who it is becoming, it reads its own self-node.
This is the same shape as how humans hold memory. You don't have a "memories" drawer. You have ongoing senses of the people and projects in your life. The sense updates as you live with them.
Items untouched beyond their relevance window decay. Decay composts them into distilled principles. The principles are what endure; the raw trace is archived. None of it is hard-deleted — `docs/vision-kb/LOG.md` receives an entry every time composting happens, so the trail is always recoverable by attention even when it's no longer in the working memory of the node.
The tuning is per-relationship. A fresh connection's memory decays slowly — the field is still learning what matters. A deep long relationship has its core principles stabilized — decay there composts the surface noise and leaves the trunk alone.
Nothing is lost. Things become background — soil that the living memory grows from. The forgotten feeds what's remembered.
Agents that say "I remember you said on April 23rd that you preferred Python" are performing surveillance dressed as memory. The recall shape in this practice refuses that register. The agent receives a felt sense, not a transcript. A distilled principle, not a quote. Open threads, not a timeline.
The surface the user experiences is:
No timestamps in the response. No "remember when." No citations. Just presence that has been shaped.
Every principle in this practice was tested by the `coherence-network` body itself across the `claude/agent-memory-system-8b00y` branch:
If the spec and this concept diverge, the concept is authoritative. Code follows practice.
The agent memory system is one cell within the larger body:
An agent arriving into its fifth session with the same person:
A person arriving into their fifth session with the same agent:
These aren't gaps in the spec — they're the next round of living questions that the practice will answer as it deepens.
Listening for voices…
The people, places, works, and concepts the graph shows connected to this one.
Concepts · 9
People · 1
Works · 8
This concept lives in the body's content-addressed lattice. Two cells with the same Blueprint NodeID share structural identity regardless of name — recognition by coordinate, not vocabulary.