Skip to content

v1.14.0

Choose a tag to compare

View all tags
@github-actions github-actions released this 26 Apr 18:22
· 91 commits to main since this release
v1.14.0
60a93da

Engram v1.14 — Memory Conflict Surfacing

Your agent can now detect when persistent memories contradict each other, resolve safely when confidence is high, and ask you in conversation when it is not sure.

No CLI flow to learn. No dashboard required. No breaking changes.

Memory Conflict Surfacing

This release closes a long-standing problem in persistent agent memory: what happens when memory contradicts itself.

Classic examples:

  • Three months ago you said “use Clean Architecture”; today you say “move to Hexagonal”. Without conflict awareness, an agent averages both and applies a Frankenstein architecture.
  • Two people on a shared project give contradictory context to the same memory layer.
  • You changed stack six months ago, but old memories still influence new decisions.

Before v1.14, Engram stored all of those memories and agents operated over partial truths. Now Engram detects the collision when new memory is saved.

How it works

The flow changes underneath normal memory saves:

  1. New memory is saved.
  2. Engram searches similar candidates with FTS5 plus topic-family matching.
  3. If candidates exist, the agent receives them with a judgment_id.
  4. The agent decides:
    • Same thing, no conflict → mark and continue.
    • Real contradiction with high confidence → resolve directly.
    • Not sure → ask you naturally in the conversation.
  5. The resolution is stored with a full audit trail: who decided, when, model, confidence, and reason.

When the agent asks, it is not a modal or a clickable prompt. It is normal chat:

“I see two architecture memories that conflict: one says Clean Architecture and one says Hexagonal. Which one is current?”

You answer normally. The agent persists the answer. Done.

What to expect after updating

  • Cleaner searches: superseded memories are surfaced with context such as supersedes: #42.
  • Real audit trail: each decision records who, when, model, confidence, and rationale.
  • Multi-persona ready: if two different agents judge the same pair, both judgments persist instead of overwriting each other.
  • Zero configuration: existing Engram users can upgrade and keep working.

Engram Cloud and deployment improvements

  • Added official multi-arch Docker image publishing for Engram Cloud:
    • ghcr.io/gentleman-programming/engram-cloud:latest
    • ghcr.io/gentleman-programming/engram-cloud:v1.14.0
    • ghcr.io/gentleman-programming/engram-cloud:1.14.0
    • ghcr.io/gentleman-programming/engram-cloud:1.14
    • ghcr.io/gentleman-programming/engram-cloud:1
  • Supported image platforms: linux/amd64 and linux/arm64.
  • Added a GHCR compose example for Dokploy, Coolify, Portainer, and VPS deployments.
  • The cloud image now defaults to engram cloud serve, so registry-based platforms do not need a custom command override.
  • Session deletes now propagate through sync, so cleaning empty local sessions can remove them from cloud/other clients instead of leaving stale empty sessions behind.

What is not in this release

Phase 2 will add:

  • Cloud sync for memory conflicts. Conflict data is local-only in this release.
  • Maintainer dashboard views for team disagreements.
  • Decay activation for old decisions that need ratification.
  • Semantic detection via embeddings. Phase 1 uses FTS5 plus agent judgment.

The schema already supports the roadmap; the remaining work is logic and UX.

Stats

  • 17 MCP tools, up from 16.
  • 13 tools in the agent profile, up from 12.
  • Around 50 new tests.
  • 0 known regressions.
  • 0 breaking changes.
  • Coverage: store 78.5%, MCP 94.9%.

Upgrade

brew upgrade engram

For Engram Cloud via Docker:

image: ghcr.io/gentleman-programming/engram-cloud:v1.14.0
Assets 9
Loading