<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>DEV Community: Silas Okanlawon</title>
    <description>The latest articles on DEV Community by Silas Okanlawon (@sighlars).</description>
    <link>https://dev.to/sighlars</link>
    <image>
      <url>https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4255899%2Fa0e84d31-1bf1-48f1-a760-d4d58f3c2ef3.jpg</url>
      <title>DEV Community: Silas Okanlawon</title>
      <link>https://dev.to/sighlars</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/sighlars"/>
    <language>en</language>
    <item>
      <title># I Spent 6 Months Building an AI Coding Memory Layer to Reduce Token Waste</title>
      <dc:creator>Silas Okanlawon</dc:creator>
      <pubDate>Fri, 20 Nov 2026 15:10:47 +0000</pubDate>
      <link>https://dev.to/sighlars/-i-spent-6-months-building-an-ai-coding-memory-layer-to-reduce-token-waste-4586</link>
      <guid>https://dev.to/sighlars/-i-spent-6-months-building-an-ai-coding-memory-layer-to-reduce-token-waste-4586</guid>
      <description>&lt;p&gt;AI coding agents are getting better at writing code, navigating codebases and handling increasingly complex tasks.&lt;/p&gt;

&lt;p&gt;But as projects grow, another problem becomes harder to ignore: agents repeatedly spend tokens rediscovering information they have already encountered.&lt;/p&gt;

&lt;p&gt;They explore the same files, reconstruct previous decisions and investigate problems that were already understood in earlier sessions.&lt;/p&gt;

&lt;p&gt;I spent the last six months building &lt;strong&gt;OpenLotus&lt;/strong&gt;, an independent memory layer designed to address this problem.&lt;/p&gt;

&lt;h2&gt;
  
  
  The problem with relying on conversation history
&lt;/h2&gt;

&lt;p&gt;While working on a project, an AI coding agent accumulates valuable knowledge.&lt;/p&gt;

&lt;p&gt;It learns the architecture, identifies technical constraints, makes implementation decisions, discovers failed approaches and tracks what still needs to be done.&lt;/p&gt;

&lt;p&gt;The problem is that much of this knowledge exists inside the agent's conversation.&lt;/p&gt;

&lt;p&gt;And conversations are not reliable long term storage.&lt;/p&gt;

&lt;p&gt;Consider what happens in a few common situations:&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Session resets:&lt;/strong&gt; A new session may not have access to the previous conversation or the reasoning accumulated during it.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Context compaction:&lt;/strong&gt; When a conversation approaches the context window limit, an agent may compress its history into a shorter summary. This helps it continue working, but important details can be lost in the process.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Context window pressure:&lt;/strong&gt; Repeatedly loading instructions, conversation history and previously explored information consumes tokens that could otherwise be used for implementation.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Switching agents:&lt;/strong&gt; Moving from Claude Code to OpenCode or another coding environment does not automatically transfer everything the previous agent learned about the project.&lt;/p&gt;

&lt;p&gt;This creates a familiar experience.&lt;/p&gt;

&lt;p&gt;You ask an agent to continue working, and it starts investigating something you already investigated. It revisits architectural decisions, reads previously explored files or reconstructs context from scratch.&lt;/p&gt;

&lt;p&gt;The model may be perfectly capable of understanding the project. The problem is that the knowledge it acquired has no reliable, independent place to live.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why this matters for token usage
&lt;/h2&gt;

&lt;p&gt;Repeated exploration has a cost.&lt;/p&gt;

&lt;p&gt;Every unnecessary file read, reconstructed decision, repeated investigation and replayed conversation consumes additional tokens.&lt;/p&gt;

&lt;p&gt;It also puts pressure on the available context window, leaving less room for the actual task.&lt;/p&gt;

&lt;p&gt;Research into memory and retrieval systems suggests that this problem can be addressed more efficiently than simply providing agents with more conversation history.&lt;/p&gt;

&lt;p&gt;For example, a ContextSniper study referenced in the OpenLotus research collection reports token reductions of 51.5% for OpenClaw and 38.9% for Claude Code under its tested conditions.&lt;/p&gt;

&lt;p&gt;A separate Mem0 comparison reports 166 prompt tokens for retrieval based memory compared with 594 for naive history, at equal answer quality.&lt;/p&gt;

&lt;p&gt;These are findings from external research, not OpenLotus benchmarks. I have deliberately kept that distinction clear.&lt;/p&gt;

&lt;p&gt;The underlying idea is straightforward: an agent does not always need more context. It needs access to the relevant information without repeatedly reconstructing it.&lt;/p&gt;

&lt;p&gt;Research: &lt;a href="https://www.openlotus.io/research" rel="noopener noreferrer"&gt;https://www.openlotus.io/research&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  I wanted project memory to exist independently of the agent
&lt;/h2&gt;

&lt;p&gt;Existing approaches provide useful pieces of persistent context.&lt;/p&gt;

&lt;p&gt;Instruction files such as &lt;code&gt;CLAUDE.md&lt;/code&gt; and &lt;code&gt;AGENTS.md&lt;/code&gt; help establish project rules. Built in memory systems can preserve selected information. Conversation summaries help agents continue previous work.&lt;/p&gt;

&lt;p&gt;But I wanted to address a broader problem.&lt;/p&gt;

&lt;p&gt;What if project knowledge belonged to the project itself, rather than a particular agent, conversation or development environment?&lt;/p&gt;

&lt;p&gt;A shared record that could preserve decisions, constraints, discoveries and implementation progress, and remain accessible regardless of which coding agent was being used.&lt;/p&gt;

&lt;p&gt;That idea became OpenLotus.&lt;/p&gt;

&lt;h2&gt;
  
  
  Introducing OpenLotus
&lt;/h2&gt;

&lt;p&gt;OpenLotus is an independent project memory and state layer for AI coding agents.&lt;/p&gt;

&lt;p&gt;At its centre is a shared data structure called &lt;strong&gt;ProgressMap&lt;/strong&gt;. It provides a persistent representation of a project's evolving state that compatible agents can access and update through the CLI and MCP.&lt;/p&gt;

&lt;p&gt;Instead of storing another copy of the entire conversation, OpenLotus organises project knowledge into structured information:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Plans:&lt;/strong&gt; What the project is trying to accomplish and what work remains.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Decisions:&lt;/strong&gt; Architectural choices, technical tradeoffs and their reasoning.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Constraints:&lt;/strong&gt; Requirements and boundaries future implementation must respect.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Lessons:&lt;/strong&gt; Discoveries, failed approaches and knowledge worth preserving.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Evidence and changes:&lt;/strong&gt; What has actually happened in the repository, including changes that may not yet be committed.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Drift:&lt;/strong&gt; Differences between the project's intended direction and its observed implementation state.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The goal is not to preserve every interaction.&lt;/p&gt;

&lt;p&gt;It is to preserve the information that makes future development more informed and less repetitive.&lt;/p&gt;

&lt;h2&gt;
  
  
  How OpenLotus works
&lt;/h2&gt;

&lt;p&gt;OpenLotus has three main components.&lt;/p&gt;

&lt;h3&gt;
  
  
  1. CLI and MCP
&lt;/h3&gt;

&lt;p&gt;A local repository can be paired with an OpenLotus project using:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;npx openlotus pair
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The CLI handles project pairing, while MCP gives compatible coding agents access to the shared project state.&lt;/p&gt;

&lt;p&gt;This means project memory is not tied to one coding environment. Different agents can work with the same underlying record rather than maintaining isolated versions of project knowledge.&lt;/p&gt;

&lt;h3&gt;
  
  
  2. ProgressMap
&lt;/h3&gt;

&lt;p&gt;ProgressMap is the central representation of project knowledge.&lt;/p&gt;

&lt;p&gt;It brings together plans, decisions, actions, changes, lessons and other information that describes the project's evolving state.&lt;/p&gt;

&lt;p&gt;When a new session begins, an agent can retrieve relevant project information instead of reconstructing everything from previous conversations.&lt;/p&gt;

&lt;p&gt;It also provides a basis for comparing declared project intent with observed repository reality.&lt;/p&gt;

&lt;h3&gt;
  
  
  3. Dashboard and Agent Workspace
&lt;/h3&gt;

&lt;p&gt;OpenLotus exposes the shared project record through three interfaces:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Progress Map:&lt;/strong&gt; A visual representation of project plans, decisions, changes and progress.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Dashboard:&lt;/strong&gt; A human readable overview of project activity, drift and areas requiring attention.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Agent Workspace:&lt;/strong&gt; An interface for interacting with an agent that has access to the shared project state.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The important part is that the human and the agent work from the same underlying project record.&lt;/p&gt;

&lt;h2&gt;
  
  
  What happens when context gets compacted?
&lt;/h2&gt;

&lt;p&gt;Context compaction is useful. It allows an agent to continue working when its conversation becomes too large for the available context window.&lt;/p&gt;

&lt;p&gt;However, a compressed conversation is not necessarily a complete record of the project.&lt;/p&gt;

&lt;p&gt;An architectural decision might be reduced to a short statement without its original reasoning. A failed approach might disappear from the summary. An unresolved constraint might lose the detail needed to avoid repeating an earlier mistake.&lt;/p&gt;

&lt;p&gt;OpenLotus addresses a different layer of the problem.&lt;/p&gt;

&lt;p&gt;Important decisions, constraints, lessons and project progress can be maintained independently of the conversation. When another session starts, the agent can retrieve that information again.&lt;/p&gt;

&lt;p&gt;The distinction is simple:&lt;/p&gt;

&lt;p&gt;A context window is working space.&lt;/p&gt;

&lt;p&gt;Compaction manages that working space.&lt;/p&gt;

&lt;p&gt;OpenLotus maintains the project record that future working sessions can return to.&lt;/p&gt;

&lt;p&gt;It does not eliminate context windows or replace compaction. It provides a more durable source of project knowledge alongside them.&lt;/p&gt;

&lt;h2&gt;
  
  
  What I have not claimed
&lt;/h2&gt;

&lt;p&gt;I want to be careful about token savings.&lt;/p&gt;

&lt;p&gt;External research supports the idea that structured memory and retrieval can reduce unnecessary context consumption. But I have not yet established a reliable percentage of token savings attributable specifically to OpenLotus through controlled benchmarks.&lt;/p&gt;

&lt;p&gt;That requires testing identical repositories and tasks with and without OpenLotus, while measuring token consumption and task success.&lt;/p&gt;

&lt;p&gt;Reducing token usage by making an agent perform worse would defeat the purpose.&lt;/p&gt;

&lt;p&gt;The objective is to reduce unnecessary exploration and context reconstruction without sacrificing implementation quality.&lt;/p&gt;

&lt;p&gt;That is the standard I want OpenLotus to meet.&lt;/p&gt;

&lt;h2&gt;
  
  
  Where I want to take this
&lt;/h2&gt;

&lt;p&gt;I believe AI coding infrastructure needs to evolve beyond individual agent sessions.&lt;/p&gt;

&lt;p&gt;As developers work with multiple agents, longer running projects and increasingly autonomous workflows, the project itself needs a durable source of truth.&lt;/p&gt;

&lt;p&gt;Git records changes to code.&lt;/p&gt;

&lt;p&gt;OpenLotus aims to preserve the context behind those changes: why decisions were made, what constraints exist, what has been learned and where implementation stands.&lt;/p&gt;

&lt;p&gt;That is what I have spent the last six months building.&lt;/p&gt;

&lt;p&gt;You can explore OpenLotus at &lt;a href="https://www.openlotus.io/" rel="noopener noreferrer"&gt;https://www.openlotus.io/&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;It currently supports Claude Code, OpenCode, Codex, Cursor and other MCP compatible clients.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;I would like to hear how other developers approach this problem.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;How do you preserve project knowledge across context compaction and session resets?&lt;/p&gt;

&lt;p&gt;And if you work with multiple coding agents, how do you keep them aligned on the same decisions?&lt;/p&gt;

&lt;p&gt;I am particularly interested in the technical approaches people are using.&lt;/p&gt;

</description>
      <category>agents</category>
      <category>ai</category>
      <category>coding</category>
      <category>llm</category>
    </item>
    <item>
      <title>I've spent six months building OpenLotus, an independent memory layer for AI coding agents. It preserves project decisions, context and progress across sessions and agents, reducing repeated exploration, wasted tokens and costly rework.</title>
      <dc:creator>Silas Okanlawon</dc:creator>
      <pubDate>Fri, 20 Nov 2026 14:38:23 +0000</pubDate>
      <link>https://dev.to/sighlars/ive-spent-six-months-building-openlotus-an-independent-memory-layer-for-ai-coding-agents-it-1i6n</link>
      <guid>https://dev.to/sighlars/ive-spent-six-months-building-openlotus-an-independent-memory-layer-for-ai-coding-agents-it-1i6n</guid>
      <description></description>
      <category>agents</category>
      <category>ai</category>
      <category>opensource</category>
      <category>tools</category>
    </item>
  </channel>
</rss>
