Most memory tools stop at retrieval. They can find a note if the agent asks the right question. That is not enough for coding agents. A useful agent needs the decision, the correction, the related file, the stale-warning, and the reason those facts belong together before it edits code.
That is what this release adds. OMEGA now assembles working context, not just search results. Core gives every developer context packets. Pro adds the repair loop that teaches OMEGA which memory chains were missing so the graph improves over time.
What shipped
Core: task-aware context packets and Pro compatibility modules. Pro: apply-capable memory chain backfill and context packet maintenance. Packaging: public core and licensed Pro are aligned at 1.5.4.
Context packets
Agents get a compact working set for the task: relevant memories, graph chains, warnings, and provenance under a token budget.
Memory chains
Related decisions, corrections, files, and warnings travel together, so the model gets the story instead of fragments.
Graph repair
When evaluation finds a missing link, Pro can add typed edges so future sessions retrieve the chain automatically.
1. Context packets replace raw memory dumps
Core now exposes context_packet by default. It returns a compact prompt block assembled from the current task, relevant memories, graph relationships, warnings, and provenance.
The benefit is immediate: less context stuffing, fewer repeated searches, and fewer moments where the agent misses a decision your team already made. The agent sees a working set, not a pile of search results.
2. Memory chains make recall useful
A vector hit is usually one node. A useful working memory is often a chain: the original incident, the failed approach, the final decision, the file that changed, and the later warning that an older memory is stale.
Chains make memory cheaper and more reliable. Instead of broad retrieval every time, OMEGA can follow the graph: this decision superseded that one, this lesson belongs to this subsystem, this warning should travel with this recommendation. The model gets fewer fragments and more usable context.
3. Pro closes the loop with graph repair
Core can show you where context quality is weak. Pro can fix it. The apply-capable maintenance loop evaluates packet misses and writes typed backfill edges into the graph.
That means the same missing relationship should not keep wasting future sessions. Once OMEGA learns that two memories belong together, future packets can recover the chain directly.
| Capability | Who gets it | Why it matters |
|---|---|---|
| context_packet | Core + Pro | Compact working context for the task |
| eval-context-packet | Core + Pro | Measure whether packets include the right memories |
| diagnose-context-packet | Core + Pro | Find why context assembly missed |
| backfill-context-packet dry-run | Core + Pro | Preview graph repairs before writing |
| backfill-context-packet --apply | Pro | Write typed repair edges into the graph |
| maintain-context-packet | Pro | Run eval plus capped graph repair as one loop |
Why this is more than search
Search answers the question you remembered to ask. A context engine prepares the agent for the task before it acts. It scopes by project, respects sensitivity, filters superseded memories, surfaces contradictions as warnings, and packs the result into a token budget the model can actually use.
That is the product goal behind this release: developers should feel OMEGA when the agent avoids a known dead end, cites the latest decision instead of the older one, and starts work with the same project-specific context a senior engineer would keep in their head.
Core and Pro are aligned again
We also fixed the packaging problem that blocked two Pro surfaces. The public core package now includes the modules Pro 1.5.x expects:
omega.bridge._core
omega.bridge._ingest
omega.bridge._query
omega.server.pid_registryThose imports unblock scheduled Dreaming and session-based coordination for Pro users. Core is public on PyPI. Pro remains private and licensed through the Supabase release channel.
What we verified
We verified the public package from PyPI and the private Pro wheel separately. A clean install of core 1.5.4 resolves the imports Pro reported missing. A clean Pro wheel install resolves against public omega-memory==1.5.4.
That distinction matters. Core is public and reproducible from PyPI. Pro remains licensed distribution, with the latest wheel served from Supabase to authenticated Pro users.
Upgrade path
Core users can upgrade from PyPI:
pip install -U omega-memory==1.5.4Pro users should refresh through the licensed installer or dashboard download flow. The latest Pro release row points at:
omega_memory_pro-1.5.4-py3-none-any.whlIf scheduled Dreaming or coordination still fail after upgrading, send the fresh traceback plus:
pip show omega-memory omega-memory-proThe most likely remaining cause would be an environment still importing an older core package from a different Python path.
Read the implementation details in the docs, or start from the quickstart.