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 [:project/service-a :depends-on :project/service-b]
        [?d :motivated-by ?c]
        [?c :description ?reason]]

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

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
transact('[[:project/db :name "PostgreSQL"]]', reason="Initial choice")

# Changed in session 4, transaction 11
retract('[[:project/db :name "PostgreSQL"]]', reason="Switching to CockroachDB for geo-distribution")
transact('[[:project/db :name "CockroachDB"]]', reason="Switching to CockroachDB for geo-distribution")

# Later: what did we think the database was before session 4?
query("[:find ?name :as-of 10 :where [:project/db :name ?name]]")
# → "PostgreSQL"

# What do we think now?
query("[:find ?name :where [:project/db :name ?name]]")
# → "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 :project/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. A single binary and a file. 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

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). Codex's MCP + hook wiring is manual — merge hooks/codex.toml into your config.toml.

Quick Start

from minigraf import query, transact

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

# Query stored descriptions
result = query("[:find ?d :where [?e :description ?d]]")

Storage Location

Default: memory.graph in the current working directory.

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

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. Extracts candidate entity tokens from the user's message (stop-word filtered, minimum 4 characters).
2. Queries the graph for facts whose values contain those tokens, using :valid-at set to the current UTC timestamp so only currently-valid facts are returned.
3. Falls back to a broad scan (capped by MINIGRAF_PREPARE_SCAN_LIMIT, default 50 rows) when no entity-specific results are found.
4. Returns the results as additionalContext prepended to the agent's working context for that turn.

For messages containing temporal signals (e.g. "before", "last week", "as of") with an explicit ISO date, :valid-at is set to that date instead (midnight UTC), enabling point-in-time recall.

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:

Configuration

Files

Tools

• minigraf_query — Query memory with Datalog
• minigraf_transact — Store facts (reason required)
• minigraf_retract — Retract facts (original stays in history)
• 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 progress of a running git ingestion; reports wall-clock time and final commit hash of the last completed run (including hook-driven ingestion)

Query Examples

# Basic query
query("[:find ?x :where [?e :attr ?x]]")

# Temporal query (state at transaction N)
query("[:find ?x :as-of 5 :where [?e :attr ?x]]")

# Aggregation
query("[:find (count ?e) :where [?e :description ?d]]")

# Single-hop graph traversal — what does api-gateway call?
query("[:find ?name :where [:project/api-gateway :calls ?svc] [?svc :name ?name]]")

# Two-hop join — transitive impact: what depends on key-store (directly or via one intermediate)?
query("""[:find ?svc
          :where [?mid :depends-on :project/key-store]
                 [?svc :depends-on ?mid]]""")

# Decision traceability — why did we choose asyncio?
query("[:find ?reason :where [:decision/asyncio-choice :motivated-by ?c] [?c :description ?reason]]")

# Typed entity query — list all stored components
query("[:find ?name :where [?e :entity-type :type/component] [?e :name ?name]]")

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.

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, finds a stored no-mocks preference, and writes a test using real database connections.

The transitive impact eval is the most discriminating for graph traversal: given "key-store is being replaced — what breaks?" the skill executes a 2-hop Datalog join and returns a full impact chain; without it, the agent correctly admits it cannot name the affected services.

See evals/benchmark.md for full results and per-eval breakdowns.

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)