Graph-based memory service for AI agents with semantic search, causal reasoning, and multi-hop traversal. Exposes memory as MCP tools for agent integration.
ReasonGraph exposes 13 tools with complete input schemas and descriptions. Naming is verb-forward and clear (push_memory, query_memory, trace_memory, what_if_memory). Descriptions are detailed (100 - 250 chars) and explain WHEN to use each tool. However, output schemas are NOT documented, the code shows tool registration but no return type specifications. Parameters are well-typed (string, integer, number, boolean) with descriptions, but some lack constraints (e.g., hops/top_k/max_depth have defaults but no min/max bounds stated). Error handling is absent from tool definitions, no guidance on what failures look like or how to recover. Security: no secrets in params (good), but no permission gates or audit trail declarations visible. Composition is strong: tools chain well (query_memory → trace_memory → what_if_memory form a reasoning pipeline). Overall: solid definitions with good naming and parameter coverage, but missing output schemas and error recovery patterns.
Answer in logical free text synthesized from discovered facts (requires a synthesizer configured on the service).
The directed cause->effect path from the fact nearest `from_text` to the fact nearest `to_text`, hop by hop, or an empty chain if none exists.
Delete a fact by its exact text. Set `purge_orphans` true to also erase entities left dangling by the removal; entities still used elsewhere stay.
Discover how facts connect across sessions: returns connection paths, each flagging whether it was found in another agent's session. Set `isolate` true to confine the walk to `session`.
Drop facts not accessed within the service's forget window.
List all knowledge sessions currently in memory.
Output schemas not documented. Tools return results but no structured schema is visible in code, LLMs cannot plan downstream calls or extract specific fields reliably.
Numeric parameters lack explicit bounds. hops, top_k, max_depth have defaults but no min/max constraints documented. LLMs may pass invalid values (e.g., hops=1000, top_k=-1).
No error handling guidance. Tool definitions do not specify what errors are possible, whether they are retryable, or what the LLM should do next (e.g., if a session is not found, if a fact does not exist).
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 67 | 2026-07-28+ | v2 |
Audit a fact's supersession: what it replaced ('supersedes') and what has replaced it ('superseded_by'). A non-empty 'superseded_by' is why a fact is no longer surfaced by default.
Store a memory in a knowledge session; returns the extracted entities. `resolve_conflicts=true` also checks nearby facts for contradictions and retires the ones this fact supersedes.
Retrieve ranked facts. Seeds from `session`; reasoning crosses sessions unless `isolate` is true (then it stays within `session`). `recency_weight` (0-1) favours newer facts over older contradicting ones. `as_of` (ISO timestamp) returns what was current at that moment; `include_superseded` also returns facts that have since been corrected.
Like query_memory but each fact comes with a relevance score, its created_at timestamp, and its scopes -- so you can threshold on confidence, dedupe, or say when something was remembered.
Walk the causal graph from a fact. direction='effects' traces downstream impact ('what did this cause'); direction='causes' traces back to root causes ('what led to this'). Returns the origin fact, the ordered causal chain (each hop cited to its source fact), and the terminal effects / root causes.
Correct memory: add `new_text` and remove the stale `old_text`. Set `purge_orphans` true to also erase entities left dangling by the removal (right-to-be-forgotten); shared entities are kept.
Counterfactual: if the given fact were false, which downstream effects would COLLAPSE (lose all causal support) vs SURVIVE via another path. E.g. if the rainfall fact were false, hospital disruptions lose their only path. Returns the pruned fact, the causal edges it solely supported (removed), the collapsed spans each cited to a now-unsupported source fact, and the spans an alternate path rescued.
Destructive tools (delete_memory, forget_stale) lack confirmation or dry-run support. No pattern to prevent accidental data loss when agents make mistakes.
No permission gates or scope declarations. Tools do not declare required permissions (e.g., 'read:memory', 'write:memory', 'delete:memory'), limiting audit and least-privilege configuration.