Sessions Graph

Sessions Graph is the Context Graph component for session context and cross-session recall. It is the authority on (:Session) nodes in the Context Graph family. It stores free-form text assertions — called Memories — written explicitly by agents, and makes them searchable in future sessions.

Requires Memgraph ≥ 3.6 (text search is stable from that release).

Installation

pip install sessions-graph

To use with Memgraph Context Graph:

pip install sessions-graph[memgraph-context-graph]

To use session reconciliation (entity extraction from session content):

pip install sessions-graph[reconciliation]

Quick start

from sessions_graph import SessionsGraph

graph = SessionsGraph()  # connects via MEMGRAPH_URL / MEMGRAPH_USER / MEMGRAPH_PASSWORD env vars
graph.setup()  # creates constraints and the text index (run once)

# Write a memory
mem = graph.save_memory(
    user_id="alice",
    content="Prefers Python over TypeScript",
    session_id="s-abc123",  # optional — links memory to a session for provenance
)

# Retrieve all memories for a user
memories = graph.get_memories("alice")

# Search memories by content (full-text, powered by Tantivy)
results = graph.search_memories("alice", "Python")

# Update or delete
graph.update_memory(mem.memory_id, "Prefers Python, especially for data tooling")
graph.delete_memory(mem.memory_id)

Integration with Memgraph Context Graph

Wire the SessionsGraphConnector into an AgentLink to get automatic session provenance — the connector tracks the active session_id and user_id from SessionStartEvent so you can reference them when saving memories.

from sessions_graph import SessionsGraph
from sessions_graph.connector import SessionsGraphConnector
from memgraph_context_graph import AgentLink
from memgraph_context_graph.adapters.claude import ClaudeAdapter

graph = SessionsGraph()
graph.setup()

connector = SessionsGraphConnector(graph)
link = AgentLink()
link.add_connector(connector)

adapter = ClaudeAdapter(
    link,
    session_id="s-abc123",
    session_kwargs={"user_id": "alice"},
)

# During the session, save memories via the Python API:
graph.save_memory(
    user_id=connector.active_user_id,
    content="User works primarily in the ai-toolkit repository",
    session_id=connector.active_session_id,
)

Graph schema

(:User {user_id})
    ├─[:HAS_MEMORY]──▶ (:Memory {memory_id, user_id, path, content, created_at, updated_at})
    │                          │
    │                       [:ABOUT]──▶ (:Project {key})   (files under /memories/projects/<key>/)
    │                          ▲                        │
    │            [:PRODUCED_MEMORY]              [:HAS_CHUNK]
    │                          │                        ▼
    └─[:HAD_SESSION]─▶ (:Session {session_id,   (:Chunk {hash, text})
                                  reconciliation_status,     ▲
                                  reconciled_at})  [:HAS_CHUNK]
                              │                        │
                        [:HAS_ACTION]                  │
                              ▼                        │
                        (:Action) ─────────────────────┘
                              │
                                              (:Entity)-[:MENTIONED_IN]->(:Chunk)

Every Memory is a file in the user's /memories tree: path is unique per user, and a Memory under /memories/projects/<key>/ links to that project. See Memory files.

(:User)-[:HAD_SESSION]->(:Session) is written by SessionsGraphConnector on session start; it is the join key other Context Graph components hang off of.

(:Action) is owned by Actions Graph; (:Chunk) and the extracted entity nodes are owned by unstructured2graph. See Session reconciliation below for how they get linked.

Memory files

memory_store(user_id) exposes a user's Memories as the file tree the Claude API memory tool (memory_20250818) works on: view, create, str_replace, insert, delete and rename on paths under /memories, returning the result and error strings that tool documents. Directories aren't stored; they are the path prefixes their files share.

store = graph.memory_store("alice", session_id="s-1")
store.create(
    "/memories/feedback/testing.md", "---\ndescription: Prefer real Memgraph in tests\ntype: feedback\n---\n..."
)
store.view("/memories")  # listing, two levels deep
store.execute({"command": "view", "path": "/memories/feedback/testing.md"})  # a tool call's input as-is

• Paths. Must start with /memories. ./.. segments, backslashes, control characters and percent-encoded dots or slashes are rejected, and repeated slashes collapse.
• Writes. create overwrites an existing file. An edit (str_replace/insert) applies only if the file still holds what the edit was computed from, so a stale read from another session fails instead of clobbering. A file can't be empty or larger than 100 KB.
• Views. A file view stops near 16,000 characters and says how to page on with view_range.
• History. Every write that discards text keeps it first as a (:MemoryVersion), created in the same query as the write. That covers an overwriting create, str_replace, insert, and delete (marked deleted). The newest 20 per file are kept. store.versions(path) lists them, newest first; a live file's history follows it across renames, and a deleted file's history stays under its last path.
• Provenance and projects. With session_id, every write records (:Session)-[:PRODUCED_MEMORY]->(:Memory). A file under /memories/projects/<key>/ is linked -[:ABOUT]->(:Project {key}), and a rename re-links it.

Apps built on the Claude API get the same files through the SDK's tool runner (pip install 'sessions-graph[anthropic]'). The application supplies the user; the model never does.

from sessions_graph.anthropic_memory import MemgraphMemoryTool  # AsyncMemgraphMemoryTool for the async client

memory = MemgraphMemoryTool(graph, user_id="alice", session_id="request-42")
client.beta.messages.tool_runner(model=..., max_tokens=..., tools=[memory], messages=[...]).until_done()

Harnesses reach it as the memory tool, which memgraph-context-graph serves once the user makes Context Graph their memory backend; see memgraph-context-graph § Memory.

Text search

Sessions Graph uses Memgraph text search (powered by Tantivy) for search_memories. The text index is created on setup():

CREATE TEXT INDEX memory_content_index ON :Memory(content);

Searches run as:

CALL text_search.search_all('memory_content_index', 'Python')
YIELD node AS m, score
WHERE m.user_id = 'alice'
RETURN m.content, score
ORDER BY score DESC
LIMIT 10;

The query string follows Tantivy query syntax.

Session reconciliation

A session's Actions Graph content (Messages, ToolCalls, ToolResults) is mostly opaque text today. Session reconciliation runs that content through unstructured2graph's chunk + entity-extraction pipeline -- GLiNER2 over hygm's default model by default, overridable to another ExtractionBackend (e.g. LightRAG) via reconcile_session(..., extraction_backend=...) -- turning it into queryable graph entities linked back to the session that produced them — see CONTEXT.md for the Session Reconciliation / Reconcilable Content / Reconciliation Status terminology.

Memory files are reconciled on their own, not with a session, because a file outlives the session that first wrote it.

• Trigger. Every write marks the file extraction_status = 'pending'.
• What runs. reconcile_memory(memory_id) first removes what the file's previous text produced. That covers relations carrying its id as source_id, its share of MENTIONED_IN.sources, and any chunks and entities nothing else backs. It then extracts from the current text, with no LLM.
• Sweep. sessions-graph reconcile --pending processes pending files before sessions.
• Deletes. Deleting a file removes what it produced the same way.
• No double extraction. memory tool calls are left out of session content, so their text isn't extracted twice.

Reconciliation writes no session summary. Recall answers from the turns themselves and uses the extracted graph as an index into them, so a summary would be one more LLM call whose output nothing reads.

This requires the sessions-graph[reconciliation] extra and no LLM API key. A key (llm.anthropic_api_key or llm.openai_api_key) only enables ontology derivation; without one, users stay on their adopted ontology.

Reconciliation never runs inside a hook itself. Local GLiNER2 extraction can exceed the timeout that hook runtimes enforce on hook commands. Instead:

• On SESSION_END and on TURN_END, SessionsGraphConnector cheaply marks the session reconciliation_status = 'pending' — no LLM calls, safe inside the hook. Turn ends matter because several runtimes (Codex, Antigravity CLI, opencode run) never report a session end; their sessions are only ever reconciled through the --pending sweep.

• The actual reconciliation run happens out-of-band, via the CLI:

  # Reconcile one session
  sessions-graph reconcile --session s-abc123
  
  # Sweep every session still marked 'pending' (e.g. from cron)
  sessions-graph reconcile --pending --limit 50
  

  The CLI runs with enforce_ontology=True: extracted entities get real type labels (:Person, :Organization, …) gated by unstructured2graph's default ontology, and anything outside it is kept but flagged ontology_conformant = false. See entity typing.

• Or, if you want it triggered automatically without a manual/cron step, opt in to a best-effort detached background process spawned right after a session ends (never on a turn end). For hook-based runtimes, set this persistently:

  memgraph-context-graph config set reconcile.auto_reconcile true
  

  This writes [reconcile] auto_reconcile = true to ~/.config/context-graph/config.toml, which memgraph_context_graph.hooks.runner._add_sessions_graph_connector reads on every hook invocation via resolve_auto_reconcile() — no shell/session restart quirks, since hook subprocesses don't reliably inherit shell profile environment variables. Direct SDK integrations that construct SessionsGraphConnector themselves must instead pass SessionsGraphConnector(graph, auto_reconcile=True) explicitly — per ADR 0002 (config-file-only-hook-resolution), no ambient environment variable is consulted here. This is fire-and-forget — if the process dies before finishing (machine sleep, crash), the session stays pending and sessions-graph reconcile --pending is the reliable backfill.

Programmatically:

from sessions_graph import SessionsGraph

graph = SessionsGraph()
graph.setup()

summary = await graph.reconcile_session(
    "s-abc123",
    enforce_ontology=True,  # match the CLI: promote entity_type to real labels
)
print(summary.status, summary.texts_considered, summary.texts_deduped)

Label promotion is opt-in and mirrors unstructured2graph's flags: the default (enforce_ontology=False, promote_labels=False) leaves entities under the backend's workspace label (gliner2 by default) with an entity_type property only; enforce_ontology=True restricts promotion to an ontology (pass ontology_path= for a custom one); promote_labels=True promotes every entity_type with no vocabulary. See unstructured2graph § entity typing.

Extracted entities land in the backend's workspace, the same one any documents ingested via unstructured2graph with that backend use, so a person or concept mentioned both in a session and in an ingested document merges into one node. Pass entity_workspace= explicitly to reconcile_session() to isolate them instead.

Content is deduplicated by hash before extraction, so re-running a sweep over already-processed content never re-extracts it. Each reconcilable unit (a message, tool call, tool result, or memory) is truncated to MAX_RECONCILABLE_CHARS (8000) before extraction, but a chatty session still has many units, so the first run can be substantial. Consider this before enabling auto_reconcile broadly.

Ontology versions

Reconciliation extracts each session under its user's adopted ontology version, and records which one on the Session as ontology_version. A user with none is on version 0, hygm.default_model(): the fixed core (User, Person and the value types), Organization/Location/Event, and the catch-alls Topic and Artifact.

Versions are graph nodes: one (:OntologyVersion {user_id, version, status, source, derive, model, pinned, source_hash}) each, holding the model as JSON. (:User)-[:ADOPTED]-> points at the current one, NEXT chains them, and adopting a version moves ADOPTED in one transaction.

A user can supply their own schema (a ManualStrategy YAML, see hygm):

sessions-graph ontology load --file coding.yaml            # --user defaults to identity.user_id
sessions-graph ontology load --file longmemeval.yaml --derive off
sessions-graph ontology show

• --derive extend (the default) adds the fixed core to the schema and pins its types: derivation may add types beside them, never merge, rename or retire one.
• --derive off uses the schema exactly as given, and nothing is derived. That's how a benchmark stays on a fixed vocabulary.

Or set it in the config file, so it follows edits:

memgraph-context-graph config set ontology.path ~/schemas/coding.yaml
memgraph-context-graph config set ontology.derive extend

Each sessions-graph reconcile applies that file to the configured user whenever its content or derive changed since their adopted version came from it. A file that doesn't validate is reported and the adopted version kept. Removing the setting deletes nothing.

Learning the ontology from the user's sessions

sessions-graph derive grows a user's model from their own sessions. You don't normally run it by hand: sessions-graph reconcile starts it, detached, when a user's count of qualifying sessions reaches 2, 4, 8, … 128, and then every 128. A qualifying session is a reconciled one with at least 2 user turns and 1,000 characters of the user's own words.

sessions-graph derive                       # --user defaults to identity.user_id
sessions-graph derive --force --seeds 1     # now, without waiting for a milestone

One run:

1. Takes an expiring claim on the user, so only one run at a time.
2. Reads the sessions since the last adopted run, sampled down to 64.
3. Derives up to three candidates (hygm.LlmRecommendationStrategy, observed with local GLiNER2).
4. Gates each candidate against the current model. When the sampled delta contains at least 8 sessions, the gate holds out a quarter (at most 8 sessions); smaller runs gate in-sample. A candidate passes when it is no worse on catch-all share (mentions typed Topic/Artifact) and on coverage (user turns with a typed relation).

The best passing candidate becomes the next version. Merged types are then relabelled in place, and the run's delta is re-extracted when types were added. If nothing passes, the candidates are kept as rejected versions and the delta rolls into the next run.

Every version records its changelog, observation counts, retired-type pool and the gate's numbers. The LLM is Anthropic when llm.anthropic_api_key is set (default claude-sonnet-5-5), else OpenAI (default gpt-5); --model overrides it. A version with derive = off is never derived from. Without either key, reconcile never spawns derive and the adopted version stays.

Recall

recall(user_id, question) returns what one user's own past sessions hold about a question: the evidence, not an answer. The caller's model answers from it.

recalled = graph.recall("alice", "Which database did we pick for the cache?")
print(recalled.render(today="2026-10-05"))  # header with reading rules, then the rows
recalled.to_json()  # the same turns and facts as data

Five lanes, each over the user's own history only:

Lane	Finds
turns	messages nearest the question by vector
text	messages matching it by full-text search
entities	entities nearest it by vector, and each one's facts
facts	extracted facts nearest it by vector
user_facts	every fact of the relation types nearest it, read from the user's turns across all sessions, whatever its head. Types rank by name and description from the user's ontology version, and retired types stay readable

Then the turns the facts were read from are added. Turns come first, then facts, each group in time order. This is the hybrid retrieval context-graph-eval run --retrieval-strategy hybrid benchmarks, through this same code: 89/100 on LongMemEval's own judge (run g417-recall-r1, 100 questions, map #390).

Widths default to the benchmarked setup; RecallConfig.from_mapping(...) overrides them (e.g. from a config file's [recall] section). Similarity is computed exactly over the user's own vectors, which costs time linear in their history: about 0.6 s for a few hundred messages, 0.7 s at 10k, 6.7 s at 100k. Without MAGE, recall runs text search alone and says so in its result.

Harnesses reach it as the recall tool: sessions-graph registers it under memgraph-context-graph's memgraph_context_graph.tools entry point (sessions_graph.tool:RECALL), and memgraph-context-graph mcp serves it, taking the user and the [recall] overrides from the config file. See memgraph-context-graph § Recall.

With memory files, recall also searches them: the memories lane matches their passages by vector and their text by full-text search. Their passages are embedded when the file is reconciled. Matching files come first as MEMORY [path, updated date] rows, ahead of the turns, with one extra reading rule. A user with no memory files gets exactly the five conversation lanes' output, header included. That's how the benchmark, whose corpora have none, stays byte-identical.

Embeddings for recall

Recall searches a user's history by vector as well as by text, so three units get an embedding property, computed inside Memgraph by MAGE's embeddings module (no model runs on the host):

Unit	Text embedded
user and assistant messages (:Action)	text, the plain message text Actions Graph writes
entities	their text, for every entity mentioned in one of the session's chunks
extracted edges	"<head> <type> <tail>. <r.text>": the fact and the sentence it was read from

Each vector also records its embedding_model. A vector from a different model counts as missing and is replaced, so vectors from two models never mix. The default model is BAAI/bge-small-en-v1.5 (384 dimensions); change it with:

memgraph-context-graph config set recall.embedding_model <huggingface-model-name>

When it runs:

• Turn end and session end: SessionsGraphConnector always spawns a detached sessions-graph embed --session <id>, whether or not auto_reconcile is on. Embedding needs no LLM, and the hook never waits on the model. Embedding at every turn end keeps recall current mid-session, and covers runtimes that never report a session end.
• Reconciliation: after extraction, the session's new entities and edges are embedded. An embedding failure never fails the reconciliation.
• Catch-up: sessions-graph embed --pending embeds every session whose embedding failed, never ran, or used another model.

The outcome is on the Session: embedding_status is completed (with embedding_model) or failed (with embedding_error).

This needs Memgraph with MAGE (memgraph/memgraph-mage), with 2 GiB of memory or more for the default model. The model downloads inside Memgraph on first use. On plain memgraph/memgraph every session records failed, and memgraph-context-graph doctor --connector sessions-graph reports the embeddings check as failing.

sessions-graph embed --session s-abc123
sessions-graph embed --pending --limit 50 --model BAAI/bge-small-en-v1.5

API reference

Method	Description
setup()	Create constraints, text index, and reconciliation indexes. Run once on first use.
drop()	Remove all Memory-related constraints and indexes.
save_memory(user_id, content, *, session_id, memory_id, path)	Persist a new Memory at path (default /memories/notes/<memory_id>.md). Returns the stored Memory object.
memory_store(user_id, *, session_id=None)	The memory tool's file commands over the user's Memories. See Memory files.
get_memories(user_id)	Return all Memories for a user, newest first.
get_memories_for_session(session_id)	Return all Memories produced by a session, newest first.
search_memories(user_id, query, *, limit=10)	Full-text search over Memory content.
update_memory(memory_id, content)	Replace the content of an existing Memory. Returns None if not found.
delete_memory(memory_id)	Remove a Memory and all its relationships.
async reconcile_session(session_id, *, extraction_backend=None, actions_graph=None, entity_workspace=None, promote_labels=False, enforce_ontology=False, ontology_path=None, embedding_model=DEFAULT_EMBEDDING_MODEL)	Run session reconciliation for one session, then embed what it wrote (see Embeddings for recall). extraction_backend overrides entity extraction to another ExtractionBackend (default: GLiNER2 over hygm.default_model(), built once per instance); lightrag_wrapper is always required regardless, since the narrative summary is always produced via its LLM. promote_labels/enforce_ontology/ontology_path control entity-type label promotion (see above). Returns a ReconciliationSummary. Requires the reconciliation extra.
get_pending_reconciliation_sessions(*, limit=100)	Return session IDs marked reconciliation_status = 'pending'.
recall(user_id, question, *, config=None, model=DEFAULT_EMBEDDING_MODEL)	What the user's own sessions hold about question, as Recalled (turns and facts; lines(), render(today), to_json()). See Recall.
embed_session(session_id, *, model=DEFAULT_EMBEDDING_MODEL)	Embed the session's messages, entities and edges that lack a vector from model, inside Memgraph. Returns counts as Embedded; records the outcome on the Session. Raises EmbeddingUnavailableError without MAGE or when the model can't load.
get_pending_embedding_sessions(*, model=DEFAULT_EMBEDDING_MODEL, limit=100)	Session IDs whose embedding failed, never ran, or used another model.