Memory OS for AI coding agents — structural code index with episodic memory
Memory OS has 19 tools with consistent naming (verb_noun pattern) and complete input schemas. However, descriptions are generic and lack LLM-optimization guidance. Most tools describe WHAT they do but not WHEN to use them or dependencies. Output schemas are undocumented, LLMs cannot plan downstream calls. Error handling is minimal (generic JSON error wrapping). Tool composition is sound (single responsibility), but parameter descriptions lack constraints (enums, ranges, formats). No tool annotations (readOnlyHint/destructiveHint) despite clear risk levels. Average description length ~120 chars (below 194 baseline); many lack actionable recovery guidance.
Add a memory entry (note, insight, or context) to the project or a specific symbol.
Show who depends on exported symbols in a file. For each exported symbol lists the unique caller files and call count. Use this before modifying a public API to understand blast radius.
Show impact of changes between two git refs or a date range.
Find callers or callees of a symbol. Shows which functions call a given symbol (callers) or which functions are called by a given symbol (callees).
Find imports that are never used.
Find circular import dependencies in a project.
Search code symbols by name across an opened project.
Output schemas undocumented. Tools return JSON but LLMs cannot infer field names, types, or pagination structure. Breaks tool chaining and forces agents to guess downstream parameter mappings.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint). Tools marked WRITE/DESTRUCTIVE in metadata but not in schema. Agents cannot distinguish safe reads from dangerous mutations.
Parameter descriptions lack constraints. 'kind' in find_symbol_tool lists examples but no enum; 'top_k' in semantic_search_tool has no min/max bounds; 'older_than_days' in memory_prune_tool unbounded. LLMs pass invalid values.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 58 | 2026-07-28+ | v2 |
Find symbols that are never called or referenced.
Composite call: returns symbol + callers + callees + memory entries + LLM summary (if cached) or generation context (if not yet generated). This is the primary tool to call before modifying a function — it replaces re-reading the whole file and its dependents.
Get the dependency graph of a project.
Retrieve memory entries for a project or symbol.
Show everything known about a file. Returns symbols, call edges, and imports for the given file.
List all indexed files in a project.
List all symbols in a file or project.
Remove old memory entries.
Search memory entries by content.
Open a project by path. Builds the index if none exists yet. Must be called before any other tool that queries a project. Multiple projects can be opened and queried in the same session.
Analyse what would break if a symbol were renamed. Returns the symbol definition, all callers, type references (textual), and import references pointing to the symbol's file.
Search for symbols using semantic similarity.
Error handling returns generic JSON wrapping. No recovery guidance, no categorization (retryable vs fatal), no invalid-value echoing. Agents cannot self-correct or decide next steps.
Descriptions lack WHEN-to-use guidance. 'Search code symbols by name' does not explain when to call find_symbol_tool vs semantic_search_tool, or what happens if symbol not found. LLMs waste reasoning cycles.