codex-agent-mem

Support Matrix

Runtime and platform support

Area Status Notes
Windows local development Supported Primary practical target today
Linux CI Supported Covered in GitHub Actions
Windows CI Supported Covered in GitHub Actions
macOS local development Expected but not verified No CI coverage yet
Codex notify capture Supported Current primary capture path
AGENTS-based continuity reinjection Supported Opt-in through --sync-project-doc when the generated pack is smaller than source context
Operational-state carry-forward Supported Objective, pending work, blockers, and guardrails are derived and injected into the working-memory pack
False-completion guardrails Supported The generated pack explicitly tells Codex not to declare completion while pending work remains
MCP stdio retrieval Supported Current primary retrieval path; one process per host connection is normal
Codex CLI / codex exec via MCP stdio Validated Short-lived Codex CLI execution uses the same local MCP server and config style as Desktop; validated separately from the long-lived Desktop host lifecycle
Codex Desktop via MCP stdio Validated Public v1.0 verification uses Codex Desktop with GPT-5.4 in a Codex environment; this is not a ChatGPT web/app connector validation
Gemini CLI via MCP stdio Validated Live validation passed with standard, read-only, and compact; mem_search returned object-root {items, count}
Claude Code via MCP stdio Validated Live validation passed with Opus 4.7. For lower-overhead workflows, avoid running multiple memory MCP layers at the same time
Claude Code with claude-mem also enabled Compatible, higher overhead observed Coexistence works, but local validation showed larger tool surface, session-start memory injection, and stop-hook latency from the additional memory layer
ChatGPT web/app connector Not validated in v1.0 Do not infer ChatGPT connector support from the Codex Desktop + GPT-5.4 validation. ChatGPT web/app needs separate connector/bridge validation before being claimed
Claude web / claude.ai Not validated in v1.0 Do not infer Claude web support from Claude Code validation. Claude Code runs local MCP stdio servers; Claude web / claude.ai needs a separate bridge or native connector path
Qwen Code via MCP stdio Validated Live local validation passed with Qwen Code 0.15.0, Ollama, qwen3.6:latest, standard, read-only, and compact; validated not_modified and clean stdio exits
Qwen local models through Ollama Validated smoke qwen3.6:35b-a3b-q8_0 and qwen3.5:9b both invoked mem_health_runtime through Qwen Code + MCP stdio; deepseek-r1:latest did not pass this path because Ollama reports no tool support for that model
DeepSeek-V3.2 through Ollama Cloud Validated cloud-backed deepseek-v3.2:cloud invoked mem_context_pack, mem_search, and mem_health_runtime through Qwen Code + MCP stdio; validated not_modified, but this is not local inference
Minimax M2.5 through Ollama Cloud Validated cloud-backed minimax-m2.5:cloud invoked mem_context_pack, mem_search, and mem_health_runtime through Qwen Code + MCP stdio; validated not_modified, but this is not local inference
Kimi Code CLI via MCP stdio MCP-connected Kimi Code CLI 1.38.0 connects to codex-agent-mem through stdio and lists 17 tools with standard, read-only, and compact; full Kimi K2.5 / Kimi K2.6 model tool-call validation is not claimed yet
GLM-5, Kimi K2.5, and Kimi K2.6 through Ollama Cloud Continuous evaluation glm-5:cloud, kimi-k2.5:cloud, and kimi-k2.6:cloud are tracked as tool-capable Ollama Cloud candidates for future live measurements
Ollama models with tool support Candidate pattern Tool-capable Ollama models can use the same MCP pattern through an MCP-capable runtime such as Qwen Code; the table records pairs already measured live
Grok / xAI Externally audited Protocol-compatible through an MCP stdio-capable orchestrator or thin JSON-RPC stdio wrapper; no local Grok CLI validation on this machine
Codex Desktop long-lived host note Documented codex-agent-mem hardens its own stdio runtime, but long-lived Desktop lifecycle issues can still amplify process accumulation; see the dedicated Codex Desktop note
mem_context_pack compact retrieval Supported Returns the compressed continuity pack and approximate token budget
mem_context_pack auto budget Supported Selects the smallest fitting budget profile from micro, normal, and full
mem_recent_changes delta retrieval Supported Summarizes what changed since the last meaningful continuity baseline
mem_scope_guard continuity guard Supported Exposes must-not-drop scope, active constraints, and closure conflicts
mem_provenance audit retrieval Supported Shows where one derived observation came from, including turn/session context and source payload hash
mem_health diagnostics Supported Reports project health score, sync quality, and structural warnings without mutating memory
mem_health_runtime diagnostics Supported Reports PID, uptime, request counts, idle timeout, and basic process lifecycle state for the active stdio server
Snapshot list/create/restore Supported Creates versioned project snapshots and restores them as derived state, not silent history mutation
FastAPI inspection API Supported Local inspection surface
Context sync metrics Supported Pack sync/skip events are stored per project and exposed through API/UI, including budget reason and build timing
Optional HTTP notify wrapper Supported Secondary path only
Codex hooks adapter Not yet supported Deferred
Codex App Server ingestion Not yet supported Deferred
Ollama adapter Not yet supported Deferred
Vector search / embeddings Not yet supported Deferred
Local inspection UI Supported Served by FastAPI at /ui, including recent changes, scope guard, closure, sync metrics, provenance, health, and snapshots

Version support

Component Current expectation
Python 3.12+
SQLite Local system SQLite with FTS5 support preferred
Codex integration model notify + MCP config