Short, mid and long term
Recent working context can be reinforced and promoted into more durable memory. Canonical reference knowledge provides a stable baseline alongside the changing work.
AXIOM / INDEPENDENT AI RESEARCH
Memory that
moves with you.
Your AI service can change. Keep decisions, their sources, and open work available for the next session or team handoff.
I built Axiom to retain working context outside a particular AI service. I completed a migration to a new embedding system and a fresh Claude instance, then tested that retained memory and weighted context selection worked in the replacement setup.
Completed private migration and testing · Firsthand implementation account. Public examples use synthetic records. The context example is deterministic; the model-swap illustration is scripted.
COMPLETED MIGRATION / MY IMPLEMENTATION
A fresh instance.
A retained working history.
My original setup used Claude through AWS Bedrock, with Titan embeddings. After access was lost, I moved embedding generation to Arctic in my own infrastructure and connected a fresh Claude instance through a different service.
I completed re-embedding and tested the replacement setup. Retained memory, retrieval, and the weighting rules I used to select useful context worked after the move. I could continue from prior work rather than reconstructing the history in a new conversation.
This completed result is my firsthand account. The public examples below demonstrate specific mechanisms using synthetic data; they do not reproduce the private migration tests or establish identical answers across models.
See how memory and weighting workThree layers.
Different responsibilities.
Source memory retained context
Source records retain observations, interpretations, decisions, and their origins outside the conversational session. An assertion remains distinct from a verified fact.
Index derived
Embeddings and SQLite projections can be rebuilt from the source record as retrieval needs change.
Model interpretation
The model uses retrieved context. Its proposed interpretation remains distinct from an observation or an accepted state.
MEMORY / TIERS, WEIGHTS AND RETENTION
Keep what matters.
Bring it back when it matters.
Axiom combines retained source material with managed memory. The tier describes how knowledge is carried forward; weighting helps select which context is useful for the current task.
Weighted context selection
Relevance, confidence, accumulated evidence, source and time inform what deserves attention. These are application-level selection rules; they do not train or transfer the underlying model's neural weights.
Revision and forgetting
Memory can be reviewed, promoted, demoted or forgotten as its usefulness changes. Retaining context is paired with managing stale information and keeping interpretations open to correction.
My tested result: after re-embedding, the replacement setup retrieved retained memory and used weighted context successfully. That is the completed private implementation described here.
WHY IT MATTERS TO A COMPANY
Keep decisions and dependencies
available to the next person.
A team needs more than an earlier answer. It needs the decision, its source, the unresolved dependency, and who can act next. Axiom's approach makes that context available outside a single chat or provider connection.
Project and team handoffs
Bring prior decisions, open questions and source-linked context into a new session so the next person can investigate without starting from a blank conversation.
Provider changes and recovery
Keep retained knowledge separate from the AI connection, then rebuild the retrieval layer when the embedding setup changes. My completed migration demonstrates that recovery path in my own system.
Evidence-led operations
Connect remembered context to evidence, ownership and follow-up. People can examine where a recommendation came from before deciding what to do.
These are practical use cases for the architecture. Company-wide adoption, quantified savings and production performance at another organization are not claimed.
Discuss a context or handoff problemPUBLIC EXAMPLE / EXECUTABLE CONTEXT RESOLVER
When evidence changes,
should the conclusion?
A remembered answer can become outdated. This deterministic example keeps the earlier observation and tests whether new evidence changes, confirms, or leaves the current conclusion unknown.
Synthetic records · No live model. The controls execute predefined state rules in your browser.
T1 / HISTORICAL RECORD
A conclusion with a source.
OBS-001: Deployment is blocked while review is pending.
INF-001: Deployment cannot proceed.
Basis: OBS-001.
Keep the original record when the working conclusion changes.
T2 / UPDATED CONTEXT
OBS-002 · Review completed; deployment is now approved.
Prior conclusion needs revision
“Deployment is blocked” becomes outdated. Preserve its evidence trail and update the working conclusion.
Deployment is approved.OBS-001 → INF-001 → OBS-002
Try missing evidence next: an earlier observation can remain in the history while its current validity is unknown.
Explore how the journal relates to a model changePUBLIC ILLUSTRATION / SCRIPTED MODEL SWAP
A different model.
The same source history.
This illustration holds the journal fixed while you step through two model views. It shows the architecture's intended separation of responsibilities.
Synthetic and scripted · No model calls or embeddings run. Matching answers and citations are written into this scenario; they are not a retrieval-equivalence result.
MODEL VIEWLLM A
Illustrated retrieval view
MODEL VIEWLLM B
Illustrated retrieval view
JOURNAL / SOURCE RECORDOne fixed synthetic fixture
100 session IDs · 100 observations · 20 inferences
O-47 → I-12 → illustrative answer
120 events · journal unchanged
The fixed record
Both model views refer to the same synthetic journal. In this scripted scenario, both answers cite O-47 and I-12.
Read the two synthetic citation records
These invented records are generated from a fixed public fixture. The other entries contain generic synthetic text.
“The demo panel color is amber.”
Session S-47. An invented observation for this illustration.
“Use amber in the demo panel.”
Based on O-47. A proposed interpretation, not an accepted state transition.
WHAT YOU CAN INSPECT
Working implementation.
Open research questions.
Private implementation
Axiom retains source-linked memory outside the conversational session. Bring the question you want to dig into; I can walk you through the architecture and migration decisions. Private source data and full test artifacts are not published here.
Request an Axiom walkthroughPublic mechanisms
The context resolver executes changed, confirming, and missing-evidence rules. The model swap is a separate scripted illustration. Neither is a benchmark of the full memory system.
Further evaluation
Broader evaluation can examine retrieval quality across larger corpora, conflicting sources, longer time horizons and additional models.
Read the research statusDesign principles and portability boundaries
- The journal is the source of record; derived indexes can be rebuilt.
- Observations, model interpretations, and accepted state remain distinguishable.
- The public context example separates observations, interpretations and accepted state. Private memory curation and forgetting have their own operational rules.
The event-journal illustration explores preserving history during revision. Private memory curation may remove items; physical erasure, retention and derived-index cleanup require explicit policy.
Account export, import, and deletion workflows are not verified here. Keeping a local journal and deleting an account with a provider are separate operations.
Axiom is my implemented memory and context work. Its broader relationship to Enterprise Intelligence remains an architectural direction; this site does not represent a combined enterprise deployment.