Skip to content
aizkaizk

Choosing the right memory tool

This page explains which problem aizk solves and when a simpler tool is a better choice. The comparison is about behavior, not winners. No head-to-head benchmark supports the feature tables below. How we evaluate explains which measurements can support a public quality claim.

Exact search, semantic file search, and aizk

Section titled “Exact search, semantic file search, and aizk”

A vault means a folder of Markdown files, notes, or project documents. Three tools cover different ways of finding information in that folder.

  • rg searches exact text and file paths without an index.
  • qmd indexes local documents for keyword and semantic search. It returns matching files or passages.
  • aizk stores sources, claims, authorship, validity periods, and sharing boundaries. It returns a ranked set of evidence that an agent can use directly.

Use exact search whenever the wording or path is known. Use semantic file search when the wording is uncertain but the matching document is the desired result. Use aizk when the answer also depends on who said something, when it was true, who may read it, or how several sources fit together.

  a question
      │
      ▼
  exact wording or path known?
      │
      ├─ yes ─▶ rg      searches the files directly
      │
      └─ no  ─▶ is a matching file or passage enough?
                    │
                    ├─ yes ─▶ qmd    searches an indexed document collection
                    │
                    └─ no  ─▶ aizk   returns ranked, sourced evidence
Need Best starting point Why
exact term or path rg it searches the files directly and needs no index
related passage with different wording qmd it combines keyword and semantic document search
current project decision aizk it can rank the maintained brief above related history
belief, preference, or observation aizk it keeps the statement attached to its speaker and kind
knowledge shared by two teams aizk one item can require membership in both organizations
earlier state of a fact aizk claims retain when they were valid and when they were recorded
sourced context for an agent aizk find returns bounded evidence with source provenance

Reproducible comparison questions

Section titled “Reproducible comparison questions”

Use synthetic fixtures when comparing retrieval systems. They make the inputs public and reproducible without exposing a private corpus. Every example below is fictional and states the retrieval behavior being tested.

Question family Example
exact lookup find the release checklist by its title
current state return the maintained Atlas migration brief rather than an older journal entry
similar titles distinguish Atlas Migration from Atlas Migration Weekly Plan
shared decision find the retry decision visible to the platform team
historical state recover the policy that was valid before a rollback
multi-source evidence gather the independent findings that support a cache change

A fair comparison must score the result each tool promises. File search succeeds when it returns the right file or passage. aizk succeeds when the authorized evidence comes from the right sources and reflects the requested history. Mixing those contracts into one score would hide what each system actually did.

Published memory systems

Section titled “Published memory systems”

The systems below overlap with aizk, but none has the same boundary.

  • Zep and Graphiti organize changing facts in a temporal knowledge graph.
  • Mem0 maintains compact user memories through add, update, and delete decisions.
  • GraphRAG builds graph communities and summaries to answer questions over document collections.
  • aizk keeps original sources beside attributed and time-bounded claims, then applies database row security before retrieval.
Capability Zep and Graphiti Mem0 GraphRAG aizk
changing facts temporal graph memory replacement not its focus separate validity and recording ranges
conflicting speakers limited representation separate user memories not its focus attributed claims with statement kinds
access control handled by the application handled by the application handled by the application forced row security in the database
overlapping organizations handled by the application handled by the application handled by the application one item may require several memberships
primary retrieval unit graph facts and text compact memories community summaries original passages, claims, and optional graph projections
deployment shape memory service hosted or self-hosted service offline indexing pipeline SQL database with replaceable model services

The repository includes a GroupMemBench adapter at src/eval/groupmem.py. A publishable comparison would still need every system to receive the same histories and questions under the same answer model, judge, and resource budget. No such result is published. External benchmarks defines that work.

Why graph retrieval stays optional

Section titled “Why graph retrieval stays optional”

Building a graph does not guarantee a better answer. The study Does Memory Need Graphs finds that raw session evidence plus independent summaries, facts, and keywords is already a strong baseline. Similarity edges can add noise, and graph summaries can improve retrieval measures while making the final answer worse because the summary displaces original evidence from the prompt.

For that reason, aizk keeps passages from the original source as primary evidence. Graph facts, profiles, and summaries are replaceable aids. Each one must improve answer quality in an ablation, which means comparing the same retrieval process with that aid enabled and disabled. Retrieval results records those comparisons.

Next

Section titled “Next”