agentMemory — Hybrid Memory System for AI Coding Agents
Gives coding agents in KiloCode, Cline, and RooCode a persistent, searchable project memory that bi-directionally syncs with their markdown memory banks — cutting repeated explanations and lost context.
Local-only storage, claimed no network access, MIT license and auditable source are positives; but the extension auto-injects 'mandatory agentMemory' instructions into .clinerules/.roo/.kilocode and rewrites .vscode/settings. without explicit user confirmation, rollback or uninstall guidance — beyond least privilege, deducted.
SKILL.md setup/start steps are readable, but repository docs self-report the dashboard storage connection as unimplemented (hardcoded zeros), possibly wrong KiloCode extension ID detection, and contradictory storage paths (.agentMemory vs ./mcp-data); error handling and edge cases thin; no high score under static review.
Target scenario (persistent memory for Cline/RooCode/KiloCode) is clear with tool arg examples; but no declared non-fit boundaries, no failure conditions when the server is not running, and triggering relies on injected config files — limited evidence.
Well-layered docs (SKILL/API/ARCHITECTURE/SECURITY), clear MIT license and author contact; but no changelog or version governance, and storage defaults conflict between package. (./mcp-data) and docs (.agentMemory), leaving hidden assumptions.
The described memory read/write/search workflow offers real marginal value for AI agents, with directly usable local markdown/JSON outputs; but the repo self-describes as incomplete and there is no execution verification evidence, statically capped below 7.
Only author claims and internal debug notes (which themselves record unfinished state); no committed test results, no CI execution evidence, no third-party corroboration; marketing claims like 'zero config / eliminates hallucinations' are unverifiable from static material.
- The skill depends on an unverified background MCP server; repo docs self-report the dashboard-to-storage connection as unimplemented.
- The extension auto-injects mandatory instructions into agent rule files and rewrites .vscode/settings. without confirmation or rollback; manually review these file changes.
- Storage paths are inconsistent between .agentMemory and ./mcp-data — confirm the actual directory before use.
- Publisher is not verified by the FollowSkills registry; identity is unknown.
- This is a static source review; no code was executed.
What does this skill do, and when should you use it?
agentMemory is a local Node.js MCP server that gives AI coding agents a structured, searchable persistent memory. It maintains a JSON database in .agentMemory/ and bi-directionally syncs with the native markdown memory banks of KiloCode (.kilocode/), Cline (.clinerules/), and RooCode (.roo/). Agents use tools like memory_search and memory_write to query and record knowledge quickly instead of scanning whole files. All data stays local and git-friendly, and a visual dashboard shows memory trends. MIT-licensed and open source for self-audit.
On first run, imports existing markdown memory banks from .kilocode/, .clinerules/, or .roo/; runs an MCP server exposing memory_search (query by text/type/tags), memory_write (record decisions by key), memory_read (retrieve by key), and memory_stats (usage analytics); stores writes in the .agentMemory/ JSON database and auto-syncs them back to markdown files such as architecture.md, systemPatterns.md, techContext.md, and progress.md; the VS Code extension adds a dashboard with memory trends, agent activity, and type distribution.
- Developers using KiloCode, Cline, or RooCode who want a searchable memory bank instead of reading entire markdown files each time
- Multi-agent teams: an OAuth architecture decision recorded by one agent can be found via memory_search by another agent, keeping implementations consistent
- New developer onboarding: synced markdown files plus memory_read quickly explain existing architecture decisions
- Developers who need cross-project memory queries via the cross-project search capability
- Privacy-conscious users who want memory data local, git-committable, and offline-capable
What are this skill's strengths and limitations?
- Bi-directional sync with native KiloCode/Cline/RooCode memory banks rather than replacing them; markdown stays human-readable and git-friendly
- Structured search by type, tags, and content is far faster than scanning whole files
- 100% local, no cloud storage, offline-first, MIT open source and auditable
- Supports cross-project memory queries and provides a visual analytics dashboard
- Requires Node.js, manual compile, and running a background MCP server — some operational overhead
- The README offers no test suite, benchmarks, or evidence of community adoption; the cross-project search setup is not documented in SKILL.md
- Internal inconsistencies: the manual-install clone URL is a placeholder, and memory_list/memory_update appear in the README tool table but not in SKILL.md
- Sync targets cover only KiloCode, Cline, and RooCode memory bank structures; other agents' memory formats are outside automatic sync
How do you install this skill?
Option 1 (VS Code extension, zero config): search "agentMemory" in the VS Code Extensions marketplace, install, and reload — the extension auto-creates the MCP server config, injects memory-first instructions, and starts bi-directional sync. Option 2 (manual, as a Skill): clone the repo, run npm install and npm run compile, then start the MCP server with npm run start-server <project_id> <absolute_path_to_workspace>. SKILL.md sits at the repo root. Note: the clone URL in the manual-install docs is a placeholder (yourusername), and the exact VSIX install path is not fully verifiable from the source.
How do you use this skill?
Once the server is running, agents follow this workflow: before a task, call memory_search({ query: "authentication", type: "pattern" }) for relevant context; during a task, use memory_read({ key: "auth-v1" }) for specific memories; after a task, call memory_write({ key: "auth-v1", type: "decision", content: "...", tags: [...] }) to record new decisions; check memory_stats({}) for analytics. Writes automatically sync to the project's markdown memory bank files. You can also open the dashboard via the VS Code command palette: "agentMemory: Open Memory Dashboard".
How does this skill compare with similar options?
The README compares three approaches: native memory banks (KiloCode/Cline/RooCode markdown files — manual, no search, no analytics), standalone generic MCP memory servers (searchable but no markdown sync), and agentMemory itself (combining sync, search, analytics, and cross-project queries).