Repository navigation
Choosing memory for a software factory #939
zlareb1
started this conversation in
Show and tell
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
Choosing memory for a software factory
For a software factory where autonomous coding agents take issues through implementation and review, start with the boundary you need to persist. If your orchestration already works, I would evaluate Loom and Mem0 first as shared memory options. Move Graphiti to the front when time-dependent relationships drive retrieval; Letta Code when you also want a stateful agent harness; or Cognee when building searchable knowledge from documents, code and conversations is the main job. This is a conditional shortlist, not a measured quality ranking. I work on Loom.
The descriptions below come from the linked primary sources, not a comparative deployment. Loom is pinned to release 1.2.0; the other repositories are pinned source snapshots. Verify capabilities in the package and edition you actually install.
Match the architecture to the work
graphiti-corein your service; supply a supported graph database, LLM and embedder. Build the surrounding ingestion, access control and tenancy.Zep and Graphiti are separate choices. Zep's repository contains examples and integrations for its service, not the service implementation. Graphiti is the open-source graph framework. The Graphiti comparison describes Zep as managed or BYOC; deploying in your cloud is not the same as operating the Graphiti library. Zep repository scope, deployment distinction.
Keep evidence separate from agent belief
Give each lesson an issue reference, repository, commit, source location, author and review status. Keep Git, the issue tracker and actual test results as the evidence behind the lesson. A stored assertion or reported successful run does not establish that the implementation is correct.
Loom's released code-link worker can mark a lesson's linked code fresh, moved, changed, unresolved or missing. A changed fingerprint calls for review; it does not establish a false lesson. Policy controls what retrieval does, and optional model revision produces a suggestion rather than rewriting the memory itself. Configuration, indexing and worker progress matter. Released code-link behavior.
For multi agent memory, scope sharing through authorized principals, not a model-supplied agent label. For autonomous bot memory, put required writes, handoff acknowledgements and review gates in awaited application calls. Offering a tool to a model does not guarantee it will call it. Record separately what the application completed, what the model claimed and what an independent check observed.
Inspect application-level evidence flows
Two Loom source examples make the evidence boundary concrete. These examples were added after the 1.2.0 release used in the comparison above; inspect their pinned source and check the server and SDK requirements before running them.
caller_selectedand supportnot_evaluated. Fetching both records does not establish that the passage supports the fact.These are inspectable application examples, not a live deployment or model-behavior evaluation. They do not replace the adoption checks below.
Test ownership and lifecycle before adoption
Self-hosting storage does not establish that processing stays local: inspect the chosen LLM, embedding, telemetry and connector configuration. Review the exact edition, model licenses, database licenses and service terms. The captured core repositories use Apache-2.0 licenses; Letta Code's license additionally excludes its brand assets. An examples repository's license does not license a hosted service.
Run one controlled issue through implementer and reviewer agents before choosing:
Deletion needs a precise definition. In Loom 1.2.0, ordinary memory deletion is a tombstone with a 14-day restore window; scoped erasure has different semantics and documented exclusions, including backups and external copies. Its audit trail is an engineering primitive, not certification, and transport audit inserts can fail open. Retention and erasure, audit limits. Establish equivalent boundaries for every candidate rather than assuming that an API named delete or forget removes every copy.
These are proposed adoption checks. No comparative deployment, reviewer-quality evaluation, cost comparison or speed test is reported here.
AI assistance
Drafted with Codex. Primary-source references and claims were checked against pinned source snapshots. Adoption checks are proposed and unrun.
All reactions