---
type: "post"
id: "most-ai-chats-remember-a-little-and-forget-62a144"
slug: "most-ai-chats-remember-a-little-and-forget-62a144"
title: "Most AI chats remember a little and forget a lot, so you end up re-explaining your project every time you start a new one."
description: "Oracle’s developer blog lays out a build for a personal knowledge base that combines content, embeddings, synthesized wiki pages and agent memory in Oracle AI Database, then exposes it to AI clients through MCP."
retrieval_nugget: "Oracle’s developer blog lays out a build for a personal knowledge base that combines content, embeddings, synthesized wiki pages and agent memory in Oracle AI Database, then exposes it to AI clients through MCP. For builders, the useful idea is a portable context layer with explicit retrieval and privacy controls, rather than relying on each chat app’s isolated memory."
published_at: "2026-11-16"
updated_at: "2026-11-16"
record_date: "2026-11-16"
date_kind: "published_at"
topics: ["mcp","agent-memory","claude"]
entities: ["Oracle","Oracle AI Database","Model Context Protocol","Claude","ChatGPT","Wojtek Pluta","Linda Haviv"]
related_patterns: ["company-memory-needs-write-loops"]
novelty_label: "new_fact"
what_changed: "Oracle provides a concrete tutorial for building a portable, self-improving second brain on Oracle AI Database, with shared context exposed to AI clients through MCP."
prior_refs: []
source_url: "https://blogs.oracle.com/developers/build-a-self-improving-second-brain-on-oracle-ai-database?tcad=1"
source_urls: ["https://blogs.oracle.com/developers/build-a-self-improving-second-brain-on-oracle-ai-database?tcad=1"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:most-ai-chats-remember-a-little-and-forget-62a144"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/most-ai-chats-remember-a-little-and-forget-62a144/","markdown":"https://newruntime.com/posts/most-ai-chats-remember-a-little-and-forget-62a144.md","json":"https://newruntime.com/posts/most-ai-chats-remember-a-little-and-forget-62a144.json"}
markdown_url: "https://newruntime.com/posts/most-ai-chats-remember-a-little-and-forget-62a144.md"
---

# Most AI chats remember a little and forget a lot, so you end up re-explaining your project every time you start a new one.

## Retrieval answer

Oracle’s developer blog lays out a build for a personal knowledge base that combines content, embeddings, synthesized wiki pages and agent memory in Oracle AI Database, then exposes it to AI clients through MCP. For builders, the useful idea is a portable context layer with explicit retrieval and privacy controls, rather than relying on each chat app’s isolated memory.

## Summary
Oracle’s developer blog lays out a build for a personal knowledge base that combines content, embeddings, synthesized wiki pages and agent memory in Oracle AI Database, then exposes it to AI clients through MCP. For builders, the useful idea is a portable context layer with explicit retrieval and privacy controls, rather than relying on each chat app’s isolated memory.

## What happened
Oracle’s July 20, 2026 developer blog, written by Director of Technical Product Marketing Wojtek Pluta and AI Engineer and Developer Advocate Linda Haviv, presents a step-by-step implementation of a “second brain” for personal content and AI work. The companion project is in Oracle’s developer hub repository. The article says the build can start locally on a Mac with Homebrew and Python 3.12, and estimates about 20 minutes for the setup. Steps one through seven run locally and headlessly; the final step moves the same code to Oracle Cloud for always-on access.

The design keeps Markdown as the portable canonical copy of source material, while using Oracle AI Database for relational records, JSON documents, vector search and memory. Oracle’s JSON Relational Duality feature presents normalized relational rows as JSON documents, allowing agents to consume document-shaped data while the database maintains underlying relations. The guide loads a MiniLM ONNX embedding model into the database and generates embeddings in SQL, avoiding an external embedding API for that path. Retrieval combines vector search with keyword search and Reciprocal Rank Fusion; the article says native Oracle hybrid search is preferable for single-table production cases, while this implementation hand-builds fusion because it ranks across posts, chunks and wiki pages.

The first agent is a research agent using Claude and tools to search stored content, read posts and wiki pages, and search the live web. The tutorial distinguishes episodic, semantic, conversational and procedural memory. In the default configuration, Oracle’s OAMP package manages semantic and conversational memory with automatic extraction; the project adds episodic and procedural extensions. A separate learning-track option uses custom tables and scheduled consolidation. The article describes ingestion and consolidation as a loop: new material is loaded, embedded, classified and incorporated into the wiki, while agent runs can be recorded and distilled into durable facts.

An MCP server exposes the same knowledge base to clients including Claude, ChatGPT and open-source clients, so changing chat applications need not move the stored brain. Oracle says its Always Free offering includes two Autonomous AI Databases, each with one Oracle CPU and 20 GB, a compute VM that can host the MCP server and 10 TB of monthly outbound transfer, with no time limit. The guide emphasizes visibility scoping, excluding private material from retrieval and memory consolidation, secret scanning, least-privilege database access, authenticated MCP hosting and human approval for write tools.

## What the archive adds
New Runtime’s earlier coverage of a five-layer memory system for an assistant used by about 1,200 people described another production design for persistent context ([the earlier note](https://newruntime.com/posts/a-habr-case-study-describes-a-five-layer-cc496d/), November 8, 2026). Oracle’s guide is a new event, not evidence that its specific architecture has been adopted in comparable production settings: it is a build tutorial that makes portability, database-backed retrieval and privacy scoping concrete.

## Trend signal
This supports the direction described by [“Company memory needs write and correction loops”](https://newruntime.com/patterns/company-memory-needs-write-loops/): useful memory is more than searching a larger pile of documents; it must be updated as work happens. Oracle’s tutorial describes ingestion, wiki maintenance and the distillation of agent runs into durable facts. It is an implementation example, not independent evidence that the loop improves outcomes for users.

## Why it matters
Teams adding memory to agents often face a choice between app-specific stores and a separate retrieval stack that must be maintained alongside the source material. Oracle’s example proposes a more unified boundary: Markdown remains exportable, while database features handle relations, vector retrieval, concurrent access and policy checks; MCP lets several clients share the same context.

That structure can reduce duplicated integration work and make privacy rules more enforceable than relying on a prompt or folder convention alone. It also concentrates risk: an incorrect visibility rule or memory consolidation can affect every connected agent. The approach is most relevant where a team needs shared, durable context and already accepts Oracle as its database platform. The tutorial does not establish that a single database is cheaper, safer or more reliable than modular alternatives.

## Limits and uncertainty
Oracle presents a tutorial and its own platform capabilities, not an independently evaluated deployment. The article’s claims about a self-improving brain describe stored information and application-level adaptation; they do not demonstrate model training or guarantee better answers over time. The guide does not establish retrieval accuracy, memory error rates, latency, operating costs at scale or how well the privacy controls withstand adversarial use. Its listed free-tier quotas and setup assumptions may not reflect every region or account configuration, and teams should validate current Oracle terms before relying on them.

## Skill to build now
Take one recurring workflow and separate its source records from its derived memory. Keep an exportable canonical format, then test whether the agent can retrieve source-backed facts and distinguish them from distilled memory. Add a visibility rule before indexing private material, and require approval for any tool that writes durable memory. Oracle’s guide gives a concrete reference architecture for those experiments, even if you use different storage and agent components.

## Reader takeaway
The valuable pattern is not simply “give the agent memory.” Keep canonical content portable, make retrieval and memory updates explicit, and enforce privacy at the data boundary. Oracle’s tutorial shows one cohesive implementation on its database and MCP stack; builders should treat its self-improvement claims as a design goal to measure, not an outcome already proven.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Mcp](https://newruntime.com/topics/mcp.md): Explore the mcp topic hub.
- topic hub: [Agent Memory](https://newruntime.com/topics/agent-memory.md): Explore the agent-memory topic hub.
- topic hub: [Claude](https://newruntime.com/topics/claude.md): Explore the claude topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)
