state-trace 0.3.0

Graph-native working memory for coding agents with causal retrieval and bounded capacity.

state-trace

Graph-native working memory for coding agents: typed memories, causal retrieval, bounded capacity, and compact briefs for small models.

state-trace is a bounded working-memory layer for coding and debugging agents that need the right file, failure, and next action under tight token budgets. It is not a replacement for a general-purpose temporal knowledge graph like Graphiti — see ARCHITECTURE.md for the honest comparison.

What it is optimized for:

• artifact-first retrieval for coding agents
• current-vs-stale task state (engine.current_state(), engine.failed_hypotheses())
• compact harness-facing briefs for smaller models
• online agent loops and post-hoc trajectory ingestion
• bounded memory with decay, compression, and lifecycle retention
• MCP-mountable, local-first deployment

Headline: SWE-bench-Verified localization — n=500

The credibility benchmark. Cold-start artifact localization on the full SWE-bench-Verified test split: given only the GitHub issue text and hints (no trajectory), rank the correct patch file at 1 and at 5.

pip install -e ".[bench]"
python3 examples/swebench_verified_eval.py --limit 500 --backends no_memory bm25 state_trace graphiti

backend	n	Artifact@1	Artifact@1 CI	Artifact@5	Artifact@5 CI	AvgLatencyMs
no_memory	500	0.000	[0.000, 0.000]	0.000	[0.000, 0.000]	0.00
bm25	500	0.176	[0.144, 0.208]	0.300	[0.262, 0.338]	0.10
state_trace	500	0.254	[0.218, 0.290]	0.376	[0.336, 0.414]	15.04
graphiti	500	0.098	[0.072, 0.126]	0.216	[0.182, 0.254]	4851.46

What this says, plainly:

• state_trace leads on both Artifact@1 and Artifact@5 against every baseline, with non-overlapping 95% CIs across the board.
• vs. Graphiti: a wide, definitive gap (A@1 0.254 vs 0.098; A@5 0.376 vs 0.216). Non-overlapping CIs on both metrics. On the same input with the same deterministic embedder/reranker stub, the typed coding-agent ontology + cold-start lexical fallback localizes the right file and puts it in the top 5 meaningfully more often.
• vs. BM25: a consistent win with non-overlapping CIs on both A@1 (lower bound 0.218 > BM25 upper bound 0.208) and A@5 (lower bound 0.336 > BM25 upper bound 0.338). BM25's pure file-token lexical search is a strong baseline; state_trace's coding-agent ontology + module-to-path translation + GitHub-URL extraction beats it decisively on cold-start localization.
• Latency: state_trace retrieves in ~15ms vs BM25's ~0.1ms vs Graphiti's ~4,850ms. For per-action memory lookups in an agent loop, the ~320× delta over Graphiti compounds meaningfully over a long session.

v0.3.0 landed a module-to-path translator in retrieve_brief's lexical fallback: dotted Python module references in issue text (astropy.modeling.separable_matrix) now resolve to file path candidates (astropy/modeling/separable.py), which pushed A@1 from 0.216 → 0.254 on n=500.

Caveats

• Graphiti is run with a deterministic hash-embedder and BM25 + cosine + BFS → RRF reranker (no LLM entity extraction). That's the same simplification graphiti_head_to_head_eval.py uses for reproducibility without API keys. A full Graphiti pipeline with GPT-4-class extraction might close some of the gap, at materially higher cost per ingest.
• Cold-start localization from issue text is only one axis. Trajectory-informed retrieval (BENCHMARKS.md) is where state_trace's larger advantage lives.

Live solve-rate — n=20 with Codex CLI + swebench docker harness

Localization leads need to be converted into downstream solve wins to matter. Running the actual swebench test suite on patches Codex CLI produces with vs. without a state-trace brief:

arm	resolved	unresolved	errored	solve-rate
state_trace	7	3	10	7/20 = 35%
no_memory	7	2	11	7/20 = 35%

Same aggregate solve-rate. But the two arms solve different instances:

• Both arms solve: 5 instances (astropy-12907, -13453, -14309, -14995, -7671)
• state_trace only: 2 instances (astropy-14598, -7166)
• no_memory only: 2 instances (astropy-14508, -7336)
• Union (at least one arm solves): 9/20 = 45%

Honest read:

• At this sample size and with Codex CLI as the downstream model, state_trace's n=500 retrieval advantage does not translate into an aggregate solve-rate advantage. The file-level proxy predicted this — Codex already localizes files near-ceiling from issue text, so the retrieval win has nowhere to compound.
• state_trace does change Codex's behavior — different instances resolve under each arm. Net-zero at n=20 could be noise or could be a genuine redirect-sideways effect. Larger N (50-100) would resolve which.
• The errors are mostly patch-apply failures — Codex produces diffs with wrong line numbers or malformed hunks, and the harness rejects them before running tests. Same pattern across both arms. That's a downstream-model problem, not a memory-layer problem.
• Union of 9/20 = 45% means routing-by-oracle between the two arms would beat either arm alone by 10 points. Suggests state_trace's context is genuinely orthogonal to Codex's baseline knowledge, just not uniformly in the correct direction.

Solve-rate caveats

• n=20 is too small for confident conclusions about the direction of state_trace's effect on solve-rate. What we can say: no big win, no big loss, identical aggregate.
• This was run against the first 20 SWE-bench-Verified instances (mostly astropy). A harder subset could shift the result either way.
• Codex CLI is a substantially stronger downstream model than a raw LLM call. Results with a smaller/weaker agent (free-tier OpenRouter, a small local model) would likely show a larger gap — in one direction or the other — because retrieval-quality wins matter more when the downstream model can't compensate.
• Reproducing this: see BENCHMARKS.md for the exact harness commands.

What makes the architecture different

Typed coding-agent ontology, not generic Entity/Edge:

• Nodes: task, observation, decision, file, goal, session, command, test, symbol, patch_hunk, error_signature, episode
• Edges: patches_file, fails_in, verified_by, rejected_by, supersedes, contradicts, solves, derived_from, precedes, motivates, and more
• Intent routing: the retrieval scorer re-prioritizes edge types per query intent (locate_file, failure_analysis, history, general).

Bounded working memory as a first-class constraint:

• enforce_capacity() runs decay, compression, and summarization on every step.
• current_state(session) answers "what's live right now" directly — cheap for state-trace, expensive for a general-purpose knowledge graph.
• failed_hypotheses(session) returns invalidated, superseded, or unrecovered-error nodes — the "don't propose this again" signal.

Local-first, MCP-mountable:

• Hot graph is an in-process networkx.MultiDiGraph. Cold storage is WAL SQLite+FTS5.
• state-trace-mcp is a stdio MCP server you can mount in Claude Code / Cursor / Codex CLI.

See ARCHITECTURE.md for why these choices matter vs. Graphiti, and BENCHMARKS.md for the smaller repo-local benchmarks.

vs. Graphiti

Graphiti is the stronger general-purpose temporal knowledge graph for AI agents. state-trace is narrower: working memory for one coding/debugging session at a time. We're not claiming to replace Graphiti — we're claiming a specific lane where the tradeoffs land differently.

Each row below is a concrete, measured axis, not a vibe.

Axis	state-trace	Graphiti	Winner for coding agents
Artifact@1 on SWE-bench-Verified, n=500	0.254 [0.218, 0.290]	0.098 [0.072, 0.126]	state-trace — non-overlapping 95% CIs
Artifact@5 on SWE-bench-Verified, n=500	0.376 [0.336, 0.414]	0.216 [0.182, 0.254]	state-trace — non-overlapping 95% CIs
Per-retrieval latency (same benchmark)	15 ms	4,851 ms	state-trace — ~320× faster
Write path per agent step	Typed insert, zero LLM calls	add_episode → LLM entity extraction each step	state-trace — cheaper, deterministic, no API key
Default deploy	Pure Python + local SQLite/JSON; state-trace-mcp stdio binary	Neo4j / Kuzu / FalkorDB graph DB + embedder + LLM	state-trace — local-first, no external services
Coding-agent ontology	Typed: file, patch_hunk, error_signature, test, command, symbol, observation, decision, task, goal, session, episode	Generic EntityNode / EntityEdge / EpisodicNode	state-trace — retrieval scorer routes on these types
"What's true right now in this session?"	engine.current_state(session) — direct O(graph) query	Inferred from temporal facts via Cypher or LLM	state-trace — first-class API
"What have I already tried and rejected?"	engine.failed_hypotheses(session) — direct query returning invalid_at + superseded + unrecovered-error nodes	Has to be inferred from invalid_at + contradictions	state-trace — first-class API
Working-memory capacity bound	enforce_capacity with decay + compression + lifecycle retention. Long-horizon pressure benchmark: Artifact@1 0.771 while staying within a 96-unit budget 100% of the time	Unbounded by design; relies on the graph DB to scale	state-trace for long debugging sessions that need a memory ceiling
Small-model brief	retrieve_brief produces ~220-token structured brief (patch_file, rerun_command, tests_to_rerun, failed_attempts, recommended_actions, …) that fits a tight budget	Returns raw nodes/facts; caller compresses	state-trace — built for small-model harnesses
MCP-mountable	state-trace-mcp stdio server in the [mcp] extra — 11 tools exposed, drop into ~/.claude/settings.json	No official MCP server; library-first	state-trace — plug straight into Claude Code / Cursor / Codex / opencode
Long-lived temporal knowledge across weeks	Scoped to a session or repo namespace; no cross-namespace fact merging	First-class; bi-temporal validity, contradiction resolution, fact supersession across episodes	Graphiti
Multi-tenant SaaS scale	Single-writer process model; authoritative graph is in-process networkx	Built for it on Neo4j/Kuzu substrate	Graphiti
Cross-session learning about users / orgs / policies	Out of scope	First-class	Graphiti

When to pick which

Use state-trace when:

• Your agent is editing code in a single debugging or refactoring session.
• You talk to an MCP client (Claude Code, Cursor, Codex CLI, opencode) and want working memory without standing up a graph DB.
• Per-action latency matters — you're calling memory on every tool invocation in an agent loop.
• You run on small models where a 220-token structured brief beats a 1,000-token raw dump.
• You need "what file should I patch / what did I already try" to be a direct query, not inferred.

Use Graphiti when:

• You need a knowledge graph of facts about the world, users, or an organization that evolves across weeks.
• Multi-tenant, multi-agent shared memory is part of the design.
• You're willing to run Neo4j/Kuzu and pay the LLM-extraction cost per ingest for the ontological payoff.
• Your retrieval patterns are richer than "which file, which test, which failed hypothesis."

They solve adjacent problems. The only reason a comparison is even interesting is that both ship as "memory for AI agents" — the honest answer is they're different products that happen to live on the same shelf.

Installation

uv sync                       # or: pip install -e .
pip install -e ".[mcp]"       # stdio MCP server for Claude Code / Cursor / Codex CLI
pip install -e ".[bench]"     # graphiti-core[kuzu] + datasets (for the headline benchmark)
pip install -e ".[llm]"       # OpenAI-backed live benchmarks + LLM ingestion
pip install -e ".[adapters]"  # LangGraph / LlamaIndex adapter shims
pip install -e ".[api]"       # FastAPI app

Distribution name: state-trace. Python import path: state_trace.

Quickstart

from state_trace import MemoryEngine

engine = MemoryEngine(capacity_limit=24.0, storage_path="memory.json")

task = engine.store(
    "Fix login by tracing the refresh token path",
    {"type": "task", "session": "auth-debug", "goal": "restore login", "file": "auth.ts", "importance": 0.92},
)
engine.store(
    "Login still returns 401 after refresh token exchange",
    {"type": "observation", "session": "auth-debug", "goal": "restore login", "file": "auth.ts",
     "blocks": [task.id], "importance": 0.88},
)
engine.store(
    "Authorization header is dropped before the retry request reaches auth.ts",
    {"type": "decision", "session": "auth-debug", "goal": "restore login",
     "related_to": [task.id], "file": "auth.ts", "importance": 0.91},
)

result = engine.retrieve("Why is login still broken?", {"session": "auth-debug", "goal": "restore login"})

Current state, live hypotheses, failed attempts

The architectural wedge. These APIs return a live view of the session without re-ranking:

state = engine.current_state(session="auth-debug", goal="restore login")
# → {"active_task": ..., "latest_observation": ..., "active_files": [...], ...}

failures = engine.failed_hypotheses(session="auth-debug")
# → [{"id": ..., "reason": ["superseded"], "content": "Login still returns 401 ..."}, ...]

current_state filters out invalidated and superseded nodes; failed_hypotheses surfaces them as "do not propose again" context. A general-purpose temporal graph has to infer this from fact updates; here it's a direct query.

MCP Server

pip install -e ".[mcp]"
state-trace-mcp

Environment config:

• STATE_TRACE_STORAGE_PATH — durable path; .db/.sqlite uses the SQLite backend. Default: ~/.state-trace/memory.db.
• STATE_TRACE_NAMESPACE — default namespace (e.g. the repo slug).
• STATE_TRACE_CAPACITY_LIMIT — working-memory budget (default 256).

Tools exposed: store, retrieve, retrieve_brief, record_action, record_observation, record_test_result, ingest_agent_log_file, current_state, failed_hypotheses, list_namespaces, graph_snapshot.

Example Claude Code config (~/.claude/settings.json):

{
  "mcpServers": {
    "state-trace": {
      "command": "state-trace-mcp",
      "env": {
        "STATE_TRACE_STORAGE_PATH": "/Users/me/.state-trace/memory.db",
        "STATE_TRACE_NAMESPACE": "repo-x"
      }
    }
  }
}

Online agent loop

engine = MemoryEngine(capacity_limit=256.0)
ctx = {"session": "auth-debug", "goal": "restore login", "repo": "example/auth-service"}

engine.record_action('open "src/auth.ts"', {**ctx, "files": ["src/auth.ts"]})
engine.record_observation(
    "AttributeError: login still fails with a 401 in src/auth.ts",
    {**ctx, "files": ["src/auth.ts"], "status": "error"},
)
engine.record_action('edit "src/auth.ts"', {**ctx, "files": ["src/auth.ts"], "action_kind": "edit"})
engine.record_test_result(
    "pytest tests/test_auth.py::test_refresh_retry",
    "tests/test_auth.py::test_refresh_retry PASSED",
    {**ctx, "files": ["src/auth.ts", "tests/test_auth.py::test_refresh_retry"]},
)

brief = engine.retrieve_brief(
    "Which file should I patch and what test should I rerun?",
    {"session": "auth-debug", "goal": "restore login"},
    mode="small_model",
)

The brief fields: patch_file, rerun_command, target_files, tests_to_rerun, current_state, failed_attempts, recommended_actions, evidence, symbols, patch_hints, confidence, token_estimate.

Trajectory ingestion

engine = MemoryEngine(capacity_limit=256.0)
engine.store_agent_log_file("examples/data/agent_logs/marshmallow__marshmallow-1867.json")

Supported inputs: normalized agent_log JSON, raw SWE-agent .traj files, raw OpenHands event JSON logs.

From iso-trace (Claude Code / Cursor / Codex / opencode sessions)

If you've accumulated session history with @razroo/iso-trace, feed it directly:

# Export a session via iso-trace's CLI
npx @razroo/iso-trace export <session-id> --json --out session.json

from state_trace import MemoryEngine
from state_trace.iso_trace_adapter import ingest_iso_trace_session

engine = MemoryEngine(capacity_limit=256.0, namespace="my-repo")
ingest_iso_trace_session(engine, "session.json")

The adapter reads iso-trace's documented Session → Turn → Event[] JSON and converts it to state-trace's agent_log format — typed nodes for files, edits, tests, errors. Months of accumulated harness history become queryable working memory without re-running the agent.

Live solve-rate (next credibility step)

examples/swebench_verified_solve_rate.py scaffolds end-to-end solve-rate measurement: state-trace brief → LLM patch proposal → SWE-bench-Verified prediction JSONL. It does not run the swebench docker harness; that step is documented in the script's header.

python3 examples/swebench_verified_solve_rate.py --limit 5 --model gpt-5.1-mini --dry-run

Storage backends

MemoryEngine(storage_path=...) picks the backend from the file extension:

• .db / .sqlite / .sqlite3 — durable SQLite with WAL journal + FTS5 seed index. Recommended for long-running agent harnesses.
• any other path — JSON blob (simple, single-writer, fine for benchmarks).

See ARCHITECTURE.md for the "why networkx + SQLite, not Neo4j" explainer.

Namespaces

engine = MemoryEngine(storage_path="memory.db", namespace="payments-api")
engine.retrieve("why is login broken?")  # scoped to payments-api by default
engine.retrieve("...", include_all_namespaces=True)  # opt out

Nodes without a namespace remain visible in every view so pre-namespace data is not lost.

Framework adapters

from state_trace.adapters import StateTraceLangGraphMemory, StateTraceLlamaIndexMemory

lg_memory = StateTraceLangGraphMemory(default_session="coding-session")
li_memory = StateTraceLlamaIndexMemory(session_id="agent-session")

Neither adapter imports the host framework; they satisfy the duck-typed memory contract used by each.

FastAPI

from state_trace.api import app  # POST /store, /retrieve, /retrieve_brief, GET /graph

Pass "explain": true on retrieve to include per-node score breakdowns.

Tests

python3 -m pytest -q

Benchmarks

Full set of repo-local benchmarks and their honest caveats lives in BENCHMARKS.md. The SWE-bench-Verified row above is the only one that's at a scale worth citing externally.

Positioning

See vs. Graphiti above for the head-to-head comparison and ARCHITECTURE.md for the architecture tradeoffs in detail. tl;dr: different products, adjacent problems — state-trace owns the narrow coding-agent working-memory lane; Graphiti owns weeks-of-history temporal knowledge graphs.