Engram

Agent memory — fingerprints, recall strategies, and the live field at engram.orkestia.dev

Engram is what an agent config remembers and what a session loads into context. Lumen is telemetry. An engram is one memory: text, an importance weight, and a fingerprint. It is not a log line, not an App Data row, and not a prompt dump.

Reads and writes go through workflows and the agents HTTP surface on https://api.orkestia.dev. You watch them in the field at engram.orkestia.dev.

The live product is this store + the field. A richer retrieval engine (salience decay, typed links) is designed and not what production ranks on. The app does not animate decay.

Surfaces

Surface	URL	Auth
Field	engram.orkestia.dev	Same org session as other `*.orkestia.dev` apps
List / delete / prune	`https://api.orkestia.dev/api/app/agents/memory`	Org session or Bearer
Live feed	`GET /api/app/agents/memory/stream`	Cookie session (`EventSource` cannot send a Bearer)
Write / recall	`agents.memory-load`, `agents.memory-save`, `agents.memory-distill`, `agents.memory-prune`	Catalog / MCP
Runner elective write	`POST /api/agents/sessions/{ref}/memory-save`	Agent session token

How it runs

session launch
    if org + config memory_enabled
        recall (last_k | importance | full | pack)
        → inject a context pack into the runner

session complete | fail
    if config memory_write_enabled
        agents.memory-distill  →  agents.memory-save
            importance < 0.1 skipped
            same fingerprint → reinforce (no second memory)
            over cap → agents.memory-prune

field
    GET /api/app/agents/memory/stream

What a memory is

You see	Meaning
Text	The remembered claim. A secret here is stored.
Importance	0–1, default `0.5`. Below `0.1` is never kept.
Fingerprint	SHA-256 of normalized text. Same digest reinforces instead of duplicating.
Access count	How often this memory was recalled into a pack.
Reinforcement count	How often the same lesson was written again.
Scope	The agent config. Optionally scoped further to an end-user or an actor.

A recall records the strategy, the cue, how many memories came back, a pack id, and later an outcome (success / failed, or an explicit rating). Distill may classify a lesson as procedural / factual / negative for the model; that label is not kept on the memory.

Flags

Flag	On	Default	Effect
`memory_enabled`	Organization	on	Org-wide read gate
`memory_enabled`	Agent config	off	Recall at launch
`memory_write_enabled`	Agent config	off	Distill after complete/fail
`memory_strategy`	Agent config	`last_k`	See Write & recall
`memory_top_k`	Agent config	`10`	Cap on returned memories
`memory_max_entries_per_config`	Organization	`10000`	Triggers prune (lowest importance, then oldest)

Read and write are separate. Turning recall on does not start writing. Writing costs an LLM call per eligible session.

In this section

Write & recall

Workflows, fingerprint, pack score, who can see whose memory.

Field & feed

Live events and what the field draws.

Agents

Where the flags sit on a config.

Custody

Engram stores text your agents wrote. It does not store source, App Data, or Lumen bodies. Prefer pointers (workflow id, ticket id, Lumen group) over payloads. The browser feed does not include who owned a memory or your organization id. See Security & compliance.

Query console and SQL

query.orkestia.dev is the org-operator door — admitted SELECT, live schema, and read-only logins. End-users still never send SQL.

Write & recall

agents.memory-* workflows — distill, fingerprint, who sees whose memory, pack scoring