01
Conversation memory is difficult for teammates to inspect, correct, or govern as shared knowledge.
Doco use case
Agent memory humans can inspect and repair
Use Doco as durable agent memory when the remembered material must remain human-readable, citable to stable blocks, versioned, and safely maintainable by both people and agents. Do not use a document workspace as a substitute for model state or ephemeral scratch context.
The real failure mode
Memory without evidence becomes folklore
An answer may sound remembered while its source has changed, disappeared, or never existed. Durable memory needs an owner, an address, and a freshness boundary.
02
A vector hit can resemble the answer without proving which current paragraph supports it.
03
An agent that rewrites a whole memory file can silently replace a newer human correction.
Operating loop
Remember through maintained documents
- 01
Browse the knowledge tree and identify the document that owns the fact; do not create a duplicate memory by default.
- 02
Search, read around the matching stable block, and retain the document version with the citation.
- 03
Answer from evidence. If coverage is incomplete or a projection is stale, disclose that boundary instead of filling the gap from memory.
- 04
When maintenance is requested, patch the smallest block under the recorded version, re-read after the write, and report what changed.
Doco stores maintained external knowledge. Session state, hidden chain-of-thought, temporary plans, and secrets belong elsewhere.
Capability map
What the system must provide
| Dimension | The need | How Doco handles it |
|---|---|---|
| Source of truth | One maintained knowledge object shared with people | Collaborative documents are the authority; summaries and concepts remain derived |
| Retrieval | Find only the relevant section instead of loading everything | Tree, outline, structured full-text search, token-budgeted reads |
| Evidence | Stable source addresses and freshness signals | Document URIs, stable block IDs, source/index versions, completeness |
| Writes | Narrow, reviewable maintenance operations | Block patches, inserts, batches, comments, relationship and summary tools |
| Concurrency | Reject stale edits instead of last-write-wins | If-Match/base_version, 409 re-read, serialized writes and snapshots |
| Portability | Export and self-hosting so memory is not trapped | Markdown/ZIP/native exports plus MIT self-hosted deployment |
Good fit
Policies, decisions, runbooks, project context, research notes, and other knowledge that people must review while agents use and maintain it.
Not the right fit
Hidden model memory, high-frequency event logs, raw telemetry, transient reasoning, credentials, or facts whose authoritative source already belongs in code or a transactional database.
Bottom line
Use the right memory
Good agent memory is not invisible recollection. It is maintained knowledge with a stable citation, an explicit version, and a human who can see and repair it.
01Is Doco a vector database?
No. Doco currently provides structured full-text retrieval and evidence-bearing document operations. It does not market that as omniscient semantic memory.
02Should an agent save every conversation to Doco?
No. Save information only when it has durable value, a clear owning document, and a reviewable update. Conversation exhaust creates noise rather than memory.
03How does an agent avoid overwriting corrected memory?
It reads the current version, writes with that version, and re-reads on a 409 conflict before merging and retrying.