Vlk: MemAct for the IDE – persistent working memory agents can prune themselves

Vlk, a native MCP server, provides persistent working memory for IDE coding agents by exposing a single tool, vlk_time_travel, that prunes dead memory slots from SQLite and injects learned lessons. This enables agents to autonomously curate their context, avoid failure loops, and retain knowledge across sessions, implementing the MemAct approach from Zhang et al. 2025.

Dead context kills long-running agents. Vlk gives them a scalpel, not a sledgehammer — and doubles as persistent memory for your IDE's coding agent. Vlk is a native MCP https://modelcontextprotocol.io server that acts as persistent working memory for IDE coding agents . It exposes a single tool — vlk time travel — that your agent calls when it detects a failure loop or context bloat. Vlk atomically prunes dead memory slots from SQLite and injects the lesson learned. Unlike ephemeral chat context that vanishes between sessions, Vlk's SQLite-backed agent history table survives restarts. Your Zed or Cursor agent picks up right where it left off — lessons intact, dead ends gone. mem id:5 London: API error 503 mem id:6 London: API error 503 retry mem id:7 London: API error 503 retry Agent calls → vlk time travel 5,6,7 , "London API down, use cached 12°C" Result: slots 5-7 deleted. Lesson persisted. Context clean. Agent unblocked. No external controllers, no fixed heuristics. The agent curates its own working memory at runtime. This is MemAct Zhang et al. 2025 https://arxiv.org/abs/2510.12635 . Zed / Cursor / Claude Desktop │ agent calls vlk time travel via stdio JSON-RPC ▼ vlk-core Rust ← persistent memory layer for the IDE │ tools/list → vlk time travel │ tools/call → atomic DELETE + INSERT ▼ SQLite WAL — agent history | Scenario | Agent behavior | |---|---| API retry storm | Agent hits same endpoint 6x with 503. Detects loop, prunes failed attempts, injects "use fallback". Continues without wasting context. | Contradictory tool output | Two search calls return conflicting facts. Agent prunes the stale one, keeps the verified source, notes the resolution. | Context window pressure | Long-running task accumulating dead branches. Agent periodically prunes abandoned reasoning paths, reclaims tokens. | Multi-turn task decomposition | Agent explores 5 approaches, 4 dead-end. Prunes dead ends, keeps winning strategy + rationale for downstream steps. | Self-correction | Agent realizes early assumption was wrong. Prunes reasoning built on it, injects corrected premise, re-derives from clean state. | IDE session persistence | You close Zed, reopen tomorrow. Agent reads agent history , sees pruned failures + injected lessons from yesterday. Continues without repeating the same mistakes. | cd vlk-core && cargo build --release Your IDE's agent will use the SQLite database vlk.db , created automatically on first run as persistent memory across sessions. // .zed/settings.json { "context servers": { "vlk": { "command": "/Users/jhonny/lab/agora/vlk-core/target/debug/vlk-core", "args": , "env": { "DATABASE URL": "sqlite:/Users/jhonny/lab/agora/vlk-core/vlk.db?mode=rwc", }, }, }, } Same pattern — point command at the binary. See MCP docs https://modelcontextprotocol.io/docs . echo '{"jsonrpc":"2.0","method":"tools/list","id":1}' | ./target/release/vlk-core → vlk time travel listed echo '{"jsonrpc":"2.0","method":"tools/call","params":{"name":"vlk time travel","arguments":{"target mem ids": 1 ,"learning":"test"}},"id":2}' | ./target/release/vlk-core → "1 slots pruned, N tokens saved" { "name": "vlk time travel", "inputSchema": { "properties": { "target mem ids": { "type": "array", "items": { "type": "integer" } }, "learning": { "type": "string" } }, "required": "target mem ids", "learning" } } @article{zhang2025memact, title = {Memory as Action: Autonomous Context Curation for Long-Horizon Agentic Tasks}, author = {Zhang, Yuxiang and Shu, Jiangming and Ma, Ye and Lin, Xueyuan and Wu, Shangxi and Sang, Jitao}, year = {2025}, url = {https://arxiv.org/abs/2510.12635} }