Repository navigation
context-graph: doctor, setup, schema learning and the Agent SDK match the docs - #604
Merged
Merged
Conversation
Found by running every page of the AI memory docs (memgraph/documentation#1847) against 0.11.0 in a sandbox. - doctor: auto_reconcile without an LLM key no longer fails the config check; reconciliation extracts locally, the key only learns the schema. - doctor: the mcp check names only the tools the config turns on, so memory shows up only once it is the memory. - setup/update/join: reinstalling the tool keeps sessions-graph's reconciliation extra once installed, and puts the sessions-graph and resources-graph commands on PATH. Turning on resources-graph used to uninstall GLiNER2 and the sessions-graph command, and never exposed resources-graph sweep. - Agent SDK: ClaudeAgentMemory.options() turns the Context Graph Claude Code plugin off through settings. With setting_sources unset the SDK loads user settings, and the plugin's hooks recorded every prompt a second time.
… anthropic 2.x Found by a second audit of the AI memory docs (memgraph/documentation#1847). - setup: a local --memgraph-url that doesn't answer (down, or refusing the login because credentials weren't set yet) is no longer swapped for setup's own container on another port; setup stops and names both ports. - sessions-graph derive: structured outputs through messages.create's output_config, which anthropic>=0.77 takes; 1.0 dropped the beta's output_format, so every candidate failed on a fresh install. derive now exits 1 when every candidate errored. - ContextGraphMemory.end_session's docstring says what ending does: embed what's left and mark the session for reconciliation.
antejavor
changed the title
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Found by running every page of the AI memory docs (memgraph/documentation#1847) against memgraph-context-graph 0.11.0 and sessions-graph 0.11.1 in a sandbox (own HOME, own Memgraph, real hook payloads, real Claude Agent SDK and OpenAI Agents SDK runs).
Fixes
doctor fails after
config set reconcile.auto_reconcile true. With no LLM key the config check failed with "reconciliation will fail", and setup ended with "1 step(s) failed". Reconciliation extracts locally with GLiNER2; the key only learns the schema. The check no longer fails, and its note says what the key is for.doctor said
mcp: serves memory, recallfor a native-memory user. It now names only the tools the config turns on.Setup's reinstall dropped local extraction. Turning on
resources-graph(Configuration page) made setup reinstall the tool, which uninstalled GLiNER2 and thesessions-graphcommand the Knowledge graph page installs, and never exposedresources-graph sweep.requirements()now keepssessions-graph[...,reconciliation]onceunstructured2graphis installed, and adds--with-executables-fromforsessions-graphandresources-graph.updateandjoinshare the path.Agent SDK apps recorded every prompt twice. With
setting_sourcesunset, the Claude Agent SDK passes no--setting-sources, so Claude Code loads user settings and the user-level Context Graph plugin's hooks fire as well.ClaudeAgentMemory.options()now setsenabledPluginsfor the plugin tofalseinsettings(flag settings outrank user settings), merging with any JSON string or settings file the app passes.Setup swapped an asked-for URL for its own container. A local
--memgraph-urlthat didn't answer (down, or refusing the login because credentials weren't set yet) fell through to setup'scontext-graph-memgraphcontainer on another port, and the config ended up pointing there. Setup now stops and names both ports; a new real-Docker test covers it.Schema learning failed on every fresh install.
derive_llmused the betaoutput_format, which anthropic 1.0 removed; a fresh install resolves anthropic 2.1.0, so every candidate failed withTypeError, andderive --forcestill exited 0. It now usesmessages.create(output_config=...)(anthropic>=0.77, floor raised, lock at 0.125.0) and exits 1 when every candidate errored. Verified with a real call on anthropic 2.1.0.Verification
./scripts/dev-memgraph.sh test memgraph-context-graph sessions-graph: 344 passed; 269 passed, 3 skipped;-m dockersetup tests 8 passed.resources-graphturned on keeps GLiNER2 and installs both commands; a real Claude Agent SDK run with the plugin enabled at user level records one user message (two before).ruff check,ruff format --checkclean.Not in this PR: every reconcile logs
Error creating entity_type typed-string constraint … already existsfromunstructured2graph— harmless, worth a separate fix.The Agent SDK change needs a release before the docs (#1847) describe it.