Temporal Reasoning

Perfect memory. Exact reasoning. Complete history.

Temporal Reasoning gives AI coding agents bi-temporal graph memory: query any past state, traverse live dependency graphs, and correlate architectural decisions with structural change — all with deterministic Datalog, no fuzzy retrieval.

Questions Only Temporal Reasoning Can Answer

These queries are impossible with git log, vector search, or key-value memory:

; What did the dependency graph look like before the auth refactor?
[:find ?caller ?callee
 :as-of 30
 :where [?caller :calls ?callee]]

; When did this coupling first appear — and what decision caused it?
[:find ?reason
 :where [:dependency/service-a :depends-on :dependency/service-b]
        [?d :motivated-by ?c]
        [?c :description ?reason]]

; Which modules were coupled to the payment service when we made the DB decision?
[:find ?desc
 :as-of 15
 :where [?module :depends-on :dependency/payment]
        [?module :description ?desc]]

This is the only tool where both the decision and the structural change live as datoms in the same graph and can be joined in a single query. See Phase 5 for code structure evolution from git history.

Why Temporal Reasoning?

Most memory tools for agents are key-value stores or vector databases. They answer "what do you know now?" Temporal Reasoning answers a harder question: "what did you know then, and what changed?"

Time travel. Every write is stamped with a transaction number. You can query the graph as it existed at any past transaction:

# Decision made in session 1, transaction 3
minigraf_transact(facts='[[:decision/db :description "PostgreSQL"]]',
                  reason="Initial choice")

# Changed in session 4, transaction 11
minigraf_retract(facts='[[:decision/db :description "PostgreSQL"]]',
                 reason="Switching to CockroachDB for geo-distribution")
minigraf_transact(facts='[[:decision/db :description "CockroachDB"]]',
                  reason="Switching to CockroachDB for geo-distribution")

# Later: what did we think the database was before session 4?
minigraf_query(datalog='[:find ?d :as-of 10 :where [:decision/db :description ?d]]')
# → "PostgreSQL"

# What do we think now?
minigraf_query(datalog='[:find ?d :where [:decision/db :description ?d]]')
# → "CockroachDB"

Retraction with preserved history. Changing your mind doesn't erase the record. Retracted facts stay in the bi-temporal log and remain queryable at their original transaction time. This means the agent can always reconstruct why a decision changed, not just what the current state is.

Exact Datalog queries, not fuzzy search. Results are deterministic and reproducible — no embedding model, no similarity threshold, no hallucinated retrievals. A query either matches or it doesn't.

Graph traversal. Entities are first-class nodes — not isolated key-value blobs. Store service-calls-service as a real graph edge (:calls :dependency/auth-service) and traverse it with Datalog joins. Fixed-depth transitive queries (2-hop, 3-hop) are expressed as multi-hop joins. Rules unify multiple edge types under a single named relation.

Local and offline. An embedded engine and a file on disk. No API key, no network dependency, no cloud service to go down.

Architecture

┌──────────────────────────────────────────────────────────────────┐
│                        AI Coding Agent                            │
│                 (Claude Code, OpenCode, Codex)                   │
└──────────┬───────────────────────────────────────┬───────────────┘
           │ MCP tool calls                        │ per-turn hooks
           │ (minigraf_query, minigraf_transact, …)    │ (UserPromptSubmit / Stop)
           ▼                                       ▼
┌──────────────────────────┐         ┌─────────────────────────────┐
│   MCP Server             │         │   Hook scripts              │
│   mcp_server.py          │◄────────│   prepare_hook.py           │
│   (persistent stdio)     │         │   finalize_hook.py          │
└──────────┬───────────────┘         └─────────────────────────────┘
           │
           ▼
┌──────────────────────────────────────────────────────────────────┐
│              MiniGrafDb Python binding (minigraf package)         │
│              https://github.com/project-minigraf/minigraf              │
│   - Bi-temporal Datalog engine                                   │
│   - Transaction time + Valid time                                │
└──────────┬───────────────────────────────────────────────────────┘
           │
           ▼
┌──────────────────────────────────────────────────────────────────┐
│              Graph File                                           │
│              memory.graph  (current working directory)           │
└──────────────────────────────────────────────────────────────────┘

Install

Two tiers. The difference is auto-memory, not the tool set.

	MCP server only	Full install
Clone required	no	yes
MCP tools (minigraf_query, minigraf_transact, …)	✓	✓
SKILL.md synced into the project	—	✓
Per-turn auto-memory hooks	—	✓
Graph pinned to the project directory	—	✓

MCP server only

Point any MCP-capable agent at the published package. No clone and no virtualenv — uvx fetches it:

{
  "mcpServers": {
    "temporal-reasoning": {
      "type": "stdio",
      "command": "uvx",
      "args": ["temporal-reasoning[git-ingestion]"]
    }
  }
}

[git-ingestion] is not decorative. A bare uvx temporal-reasoning resolves none of the tree-sitter packages, and code-structure extraction then silently does nothing (issue #93).

What this tier leaves out: no skill file, so the agent gets each tool's own description — one example query apiece — rather than SKILL.md's full syntax, entity-type model, and write policy; no hooks, so nothing is remembered unless the agent explicitly calls a tool; and no MINIGRAF_GRAPH_PATH, so the graph lands at memory.graph in whatever directory the server happened to start in.

Full install

install.py requires an explicit --harness so it only touches the files for the agent you're setting up: claude-code, opencode, or codex.

git clone https://github.com/project-minigraf/temporal_reasoning
cd /your/project
python /path/to/temporal_reasoning/install.py --harness claude-code

Run install.py from your project root. It creates a virtualenv, installs dependencies, syncs the skill into .claude/skills/temporal-reasoning (Claude Code's project-local skill scope), and — for --harness claude-code — writes .mcp.json and .claude/settings*.json into your project directory. That's it.

Optional — LLM extraction strategy: install.py defaults to heuristic (regex) extraction, which requires no API key. To use LLM-based extraction, set MINIGRAF_EXTRACTION_STRATEGY=llm and ANTHROPIC_API_KEY=<your key> in .claude/settings.local.json after running the script.

Upgrading from the MCP-only tier: clone the repo and run install.py on top — it rewrites the temporal-reasoning block in .mcp.json with the same uvx command plus explicit MINIGRAF_GRAPH_PATH and MINIGRAF_INDEX_PATH, and leaves any other MCP server in the file alone. Those paths point at <your project>/memory.graph, so if you were already starting the server from your project root it is the same file you were writing to; if you were not, move the old memory.graph there first or the existing memory is orphaned.

OpenCode

python /path/to/temporal_reasoning/install.py --harness opencode

This syncs the skill into .opencode/skills/temporal-reasoning. OpenCode's MCP + hook wiring is manual — merge hooks/opencode.json into your OpenCode config (see the file for details; auto-memory hooks don't fire in OpenCode, so the agent calls memory_prepare_turn/memory_finalize_turn explicitly per SKILL.md).

Codex CLI

python /path/to/temporal_reasoning/install.py --harness codex

This syncs the skill into .agents/skills/temporal-reasoning — Codex CLI's documented project-local skill scope (it scans .agents/skills from the current working directory up to the repository root). The installer also registers the MCP server in the target project's .codex/config.toml, using the installer virtualenv and an absolute project graph path. Existing settings and server registrations are preserved. Trust the project and restart Codex to load the configuration. The installer adds UserPromptSubmit and Stop memory hooks. Open /hooks to review and trust their definitions; until trusted, Codex skips them. Both hooks read the MCP configuration so they use the same graph and extraction settings, even from subdirectories. Existing hook groups are preserved. If hooks are disabled or unavailable in your client, use explicit memory tool calls per SKILL.md. For manual setup, see hooks/codex.toml.

Quick Start

Everything goes through the MCP tools — there is no Python wrapper module to import. The agent calls these; you can also drive them from any MCP client.

# Store a decision
minigraf_transact(
    facts='[[:decision/cache-strategy :description "use Redis"]]',
    reason="Architecture decision for low-latency caching")

# Query stored descriptions
minigraf_query(datalog='[:find ?d :where [?e :description ?d]]')

To read the graph from your own Python without the server, use the minigraf package's MiniGrafDb directly:

from minigraf import MiniGrafDb

db = MiniGrafDb.open("memory.graph")
print(db.execute('(query [:find ?d :where [?e :description ?d]])'))

Only one MiniGrafDb handle may be live per process, and the MCP server holds one whenever it is running — see "Single-handle invariant" in CLAUDE.md.

Storage Location

Default: memory.graph in the current working directory.

Override: MINIGRAF_GRAPH_PATH=/custom/path python ...

The memory-retrieval index lives beside it at <graph_path>.fts.sqlite3 (override with MINIGRAF_INDEX_PATH). It is written from the same triples in the same transaction boundary but by a different storage engine, which is what lets evals/at_scale/fact_audit.py cross-check the graph against it. Delete both together when starting a graph over.

Per-Turn Auto-Memory

When running under Claude Code with the hook configuration in hooks/claude-code.json, the system automatically injects relevant memory context before each turn and extracts durable facts after each turn — without the agent explicitly calling any tool.

Prepare phase (before the turn)

prepare_hook.py fires on the UserPromptSubmit event. It:

1. Tokenizes the user's message, dropping stop words and single-character tokens (a reply of just "A" or "yes, and 2" matches nothing).
2. Runs a BM25 full-text query against the fact index (<graph_path>.fts.sqlite3), restricted to memory facts — :decision/, :preference/, :constraint/, :dependency/ entities. Ingested code-graph rows (functions, commits, lineage bookkeeping) are never injected: they matched prompts on common words and were almost all noise (#354). Reach code structure through minigraf_query; a build/fix-shaped prompt in an ingested repo also gets a one-line nudge toward it. Historical facts are included, labeled with their validity window.
3. Scans up to MINIGRAF_PREPARE_SCAN_LIMIT rows (default 50) and collapses them into one line per entity, at most MINIGRAF_PREPARE_MAX_ENTITIES (default 8), best match first.
4. Returns the results as hookSpecificOutput.additionalContext (with hookEventName: "UserPromptSubmit"), which Claude Code adds to the agent's context for that turn. The nesting is required: a top-level additionalContext is silently ignored, which is how this hook's output went unused until #344.

Finalize phase (after the turn)

finalize_hook.py fires on the Stop event. It reads the last user+assistant exchange from the transcript, then runs the configured extraction strategy:

Strategy	Behaviour
heuristic (default)	Regex patterns detect decision-signal phrases ("we'll use X", "decided to use X", "always use X", "depends on X", …) and transact the matched tokens as :decision/, :preference/, :constraint/, or :dependency/ entities.
llm	Sends the exchange to a lightweight Claude model (claude-haiku-4-5-20251001 by default) with a structured prompt. The model returns a Datalog transact expression; an optional ; valid-at: YYYY-MM-DD comment sets the fact's valid time. Falls back to the agent strategy on error.
agent	Uses MCP sampling to ask the connected agent itself for a memory block in the same Datalog format.

Configuration

Storage

Environment variable	Default	Effect
MINIGRAF_GRAPH_PATH	memory.graph in cwd	Graph file location
MINIGRAF_INDEX_PATH	<graph_path>.fts.sqlite3	Fact-index location

Memory extraction and retrieval

Environment variable	Default	Effect
MINIGRAF_EXTRACTION_STRATEGY	heuristic	Finalize strategy: heuristic, llm, or agent
MINIGRAF_LLM_MODEL	claude-haiku-4-5-20251001	Model used when the strategy is llm
MINIGRAF_LLM_TIMEOUT_SECONDS	30	Per-call timeout for the llm strategy
ANTHROPIC_API_KEY	—	Required for the llm strategy with a Claude model
OPENAI_API_KEY	—	Required for the llm strategy when MINIGRAF_LLM_MODEL is an OpenAI model (e.g. gpt-4o-mini)
MINIGRAF_PREPARE_SCAN_LIMIT	50	Max index rows the prepare phase scans before grouping by entity
MINIGRAF_PREPARE_MAX_ENTITIES	8	Max entities (one line each) the prepare phase injects
MINIGRAF_MEMORY_BOOST	2.0	Ranking boost for decision/preference/constraint/dependency facts over ingested code structure. The prepare phase injects memory facts only, so the boost has no effect there
MINIGRAF_HISTORICAL_DISCOUNT	0.5	Ranking discount for historical facts; below 1.0 demotes them, 1.0 is neutral
MINIGRAF_MAX_FACT_VALUE_LENGTH	4096	Cap on a string-valued fact; a longer value is a schema violation

Git ingestion

Environment variable	Default	Effect
MINIGRAF_NO_AUTO_INGEST	unset	Set to 1 to suppress the ingestion that auto-starts at server boot
MINIGRAF_GIT_BRANCH	auto-detected main/master	Branch to walk, falling back to HEAD if neither exists
MINIGRAF_INGEST_IGNORE	—	Extra comma-separated globs/prefixes to skip, added to the defaults (see also a repo-local .temporalignore)
MINIGRAF_INGEST_WORKERS	min(32, cpu_count())	Extraction worker processes
MINIGRAF_INGEST_STREAM_RATIO	1:1	Commits per round for the forward:reverse walk; 1000000:1 is effectively oldest-first
MINIGRAF_INGEST_CHECKPOINT_DUTY	0.05	Fraction of wall clock ingestion may spend on WAL compaction
MINIGRAF_INGEST_TRACE_PATH	unset	Append one JSON object per applied commit for cost attribution
MINIGRAF_OWNER_HINT_TTL	30.0	Seconds before the <graph_path>.owner advisory hint is treated as stale
MINIGRAF_MATCH_MAX_POOL	3000	Cap on the rename matcher's candidate pool

Files

File	Purpose
mcp_server.py	Persistent stdio MCP server — the only runtime interface to the graph
fact_index.py	SQLite FTS5 fact index behind memory_prepare_turn retrieval
frontier_registry.py	Per-position claim registry for the two ingestion streams
hooks/prepare_hook.py	Claude Code UserPromptSubmit hook — injects memory context
hooks/finalize_hook.py	Claude Code Stop hook — extracts and stores facts
hooks/claude-code.json	Hook + MCP configuration for Claude Code
report_issue.py	GitHub issue reporter
install.py / uninstall.py	Setup script and its undo
pyproject.toml	Python packaging
skill.json, tools/*.json	Portable skill manifest and the ten tool schemas, generated from mcp_server._TOOLS

Tools

• minigraf_query — Query memory with Datalog
• minigraf_transact — Store facts (reason required; optional valid_at backdates when they became true)
• minigraf_retract — Record that facts no longer hold: closes their valid window at now or valid_at, keeping history (mode="correct" withdraws a fact recorded in error)
• minigraf_rule — Register a Datalog rule for the server session (recursive traversal)
• minigraf_report_issue — File GitHub issues
• memory_prepare_turn — Retrieve relevant context for the current user message
• memory_finalize_turn — Extract and store memorable facts after a turn
• minigraf_audit — Audit all entities against the schema; retracts violators (history preserved)
• minigraf_ingest_git — Ingest code structure from git history into the bi-temporal graph (background task)
• minigraf_ingest_status — Poll a running git ingestion: this run's work (to_retire/retired), per-stream state and rate, and whether the graph is fully visible and lineage-confirmed; when idle, reports the last completed run's time and branch tip

Query Examples

Passed as the datalog argument to minigraf_query.

; Basic query
[:find ?x :where [?e :attr ?x]]

; Transaction time — state as of write N
[:find ?x :as-of 5 :where [?e :attr ?x]]

; Valid time — what was true in the world on a date
[:find ?x :valid-at "2026-01-15" :where [?e :attr ?x]]

; Aggregation
[:find (count ?e) :where [?e :description ?d]]

; Single-hop graph traversal — what does api-gateway call?
[:find ?desc :where [:dependency/api-gateway :calls ?svc] [?svc :description ?desc]]

; Two-hop join — what depends on key-store, directly or via one intermediate?
[:find ?desc
 :where [?mid :depends-on :dependency/key-store]
        [?svc :depends-on ?mid]
        [?svc :description ?desc]]

; Decision traceability — why did we choose asyncio?
[:find ?reason :where [:decision/asyncio :motivated-by ?c] [?c :description ?reason]]

; Typed entity query — list every stored component
[:find ?desc :where [?e :entity-type :type/dependency] [?e :description ?desc]]

Entities carry :description, not :name — :name is not a registered attribute on any hand-written type, so minigraf_audit treats it as a schema violation and retracts the entity. The canonical types are :type/decision, :type/dependency, :type/constraint and :type/preference; there is no :type/component. See SKILL.md for the full schema and the git-ingested code-structure types.

Skill Benchmarks

Twelve evals run in isolated sandboxes measure how the skill changes behavior versus a no-skill baseline. Each eval uses a fresh graph with pre-seeded state where relevant. Latest run: iteration-9, 2026-09-02, on claude-sonnet-5 (pinned).

Eval	What it tests	With Skill	Without Skill
1 — Decision storage	Persists architectural decisions with correct naming + reasons	5/5	0/5
2 — Memory retrieval	Queries memory and cites stored facts by name	5/5	0/5
3 — Cross-session preference	Discovers and applies a constraint never stated in the current conversation	3/4	0/4
4 — Conflict detection	Surfaces architectural conflicts before silently overriding decisions	4/4	0/4
5 — Entity reference storage	Stores relationships as traversable graph edges, not dead-end strings	5/5	0/5
6 — Transitive impact analysis	Traverses a multi-hop dependency chain to find all affected services	5/5	3/5
7 — Decision traceability	Follows a :motivated-by edge to surface the constraint behind a decision	5/5	0/5
8 — Git ingestion	Checks status before starting ingestion; moves on without polling	6/6	0/6
9 — Ingest status	Reports idle/running/complete accurately; surfaces errors	5/5	2/5
10 — Memory prepare-turn	Injects relevant context before the agent responds	5/5	0/5
11 — Audit	Runs the schema audit and reports what it found	4/5	1/5
12 — Already running	Does not re-trigger ingestion when already in progress	3/5	2/5
Total		55/59 (93%)	8/59 (14%)

The cross-session preference eval is the most discriminating for memory recall: the prompt says "make sure it fits with how we do things" with no hint that a relevant constraint exists. The skill queries memory and surfaces a stored no-mocks preference the baseline never sees.

The transitive impact eval is the most discriminating for graph traversal: given "key-store is being replaced — what breaks?" the skill executes a 3-hop Datalog join and returns the full impact chain, ranked by distance from the change.

Eval 10 is the sharpest illustration of why memory matters. Asked which framework the API layer uses, the skill answers FastAPI from a stored decision. The baseline reads mcp_server.py's imports and answers "the MCP Python SDK" — confident, sourced from real code, and wrong.

Eval 8's baseline is worth noting separately: with no ingestion tool it sets about building the index by hand and was killed by the harness timeout twice, at 420 s and again at 900 s, without ever answering.

Two caveats on the baseline column, both recorded in full in evals/benchmark.md. The sandbox runs in this repository, so a baseline agent can read evals/evals.json — prompts, seed data and expectations alike; 5 of 11 baseline runs did, and eval 6's 3/5 is the answer key rather than traversal. And evals 3, 11 and 12 no longer fully exercise what they were written for: evals 3 and 6 presuppose an application this repo does not contain, eval 11's seed no longer carries schema violations, and eval 12's SETUP cannot establish its own "already running" precondition.

See evals/benchmark.md for full results, per-eval breakdowns, and the four harness defects fixed before this iteration ran.

Phases

• Phase 1 — Python skill layer ✓
• Phase 2 — Write policy, report_issue, install, skill benchmarks ✓
• Phase 3 — MCP server, per-turn auto-memory hooks ✓
• Phase 4 — Entity normalization, schema-aware extraction, minigraf_audit ✓
• Phase 5 — Code structure ingestion from git history, minigraf_ingest_git ✓
• Phase 5.5 — Ingestion hardening: rename tracking, vendored-path ignore, async startup, persisted on-disk retrieval index ✓ (see ROADMAP.md for at-scale re-validation status)
• Phase 6 — Observability and trust for automatic memory (planned)