OpenViking Memory Hooks — persistent memory across Codex sessions

[benediktkraus]

Built a hook system that gives Codex persistent long-term memory via OpenViking.

What it does: Every prompt gets enriched with relevant memories from previous sessions. Every session captures decisions, preferences, and context. Works across Codex, Claude Code, Gemini CLI, and OpenClaw — same memory pool.

Recall pipeline: Hybrid search (semantic + grep + RRF merge) → 5-scope system (CWD-based) → 7-dimension ranking → MMR diversity filter → temporal decay → caveman compaction → citations

Capture pipeline: Substance filter → pre-processing (strip code/tags) → content merge (>0.85 appends instead of duplicating) → auto-link (project + tool + incident relations)

Codex-specific: Installs as a local marketplace plugin with UserPromptSubmit + Stop hooks. One command setup.

git clone https://github.com/benediktkraus/openviking-hooks.git
cd openviking-hooks
./install.sh codex

Fork of Castor6/openviking-plugins with hybrid search, scoping, ranking, retry, remote mode added.

Repo: https://github.com/benediktkraus/openviking-hooks

[Mahkhmood9]

Have you noticed it preforming worse on some tasks? I personally just as happy with the openai codex --
(per-haps openviking serves as a fun project - or a tell - to extract chats from different coding agents - and for a certain type of power users)
its just a tool to extract memory/chats between different agents

https://github.com/volcengine/OpenViking/tree/4b9d17db2f05601ab0daef4ce851b0415f1ae69d#evaluation-highlights
the evaluation-highlights , dont show how it preforms with codex

maybe its more built for other products -- its actually very solid (I can said from another product i use--- codex is very good with the memory usage --- with the other big frontier model -- its not really that great its like its a real person memory codex is the only thing that truely adopts -- even in terms of token usage its very low for me atleast)

have you noticed anything weird with "openviking" ? does it does anything weird (like actually consume more tokens on certain task , and repeat itself in loops)

despite that I thikn itss a good sport and spirit , if someone is into this agentic things --- how they can be improve --- finding issues where the issue are the OEM products

its a serious thing, in the future if they make heavy api changes --- and if there is whatever subagents -- and then users go and call the api in alegitmate ways - it can poision the data (but I thikn its very improtant that there would be some users who try to modify how the coding agent -- try look up code --- try to call functions -- prompts injected into the models -- and its a thing you can modify some of the prompts sent to openai )

I honestly dont have any real problems with codex memory - and I am a very heavy user

despite it, I think for example multi agents is some thing I do not look - but its a thing -- they are prompts written to it -- and you guild the ai model with text -- and it knows to execute code or whatever -- and it can be modified to some level
https://github.com/openai/codex/blob/main/codex-rs/core/src/tools/handlers/multi_agents_spec.rs#L12-L20