CLAUDE.md Project Memory File Explained: Stop Losing AI Context Between Sessions (Lesson 1.4) — Transcript

Full transcript

1. 0:00Welcome to lesson 1.4, build persistent
2. 0:03context. This module, part of the
3. 0:06hands-on Claude code and Claude skills
4. 0:08course, is designed for software
5. 0:10engineers. We will focus on writing a
6. 0:12production-quality claude.md project
7. 0:15memory file to enhance your AI
8. 0:17workflows. The naive approach of
9. 0:19starting Claude code sessions cold leads
10. 0:21to serious drawbacks. Without context,
11. 0:24you face 20 to 40% token waste,
12. 0:26inconsistent suggestions, and silent
13. 0:29drift. The solution is to write and
14. 0:31maintain a claude.md file once to
15. 0:33provide persistent context. There are
16. 0:36three failure modes without persistent
17. 0:38context: lost context, silent drift, and
18. 0:41redundant loops, all leading to wasted
19. 0:43tokens and inconsistent code. The key is
20. 0:46that claude.md is read at session
21. 0:48initialization before your first prompt,
22. 0:51turning it into system context. This
23. 0:53diagram illustrates how Claude code
24. 0:55processes information. Your repo,
25. 0:58including claude.md, is parsed during
26. 1:00initialization to create context for the
27. 1:02Claude API. This context is combined
28. 1:05with your message for the final request.
29. 1:07The parsing order is overview, tech
30. 1:09stack, and current state. Token
31. 1:11efficiency is a major benefit. A typical
32. 1:1416K budget includes tokens for
33. 1:16claude.md, code, and your message, with
34. 1:19the rest for the response. A
35. 1:20well-written claude.md can save 20 to
36. 1:2340% of code tokens, reclaiming about
37. 1:253,200 tokens per session for actual code
38. 1:28changes. Your claude.md file should be
39. 1:31organized into three essential sections:
40. 1:34overview, tech stack, and current state.
41. 1:36These sections should be concise, around
42. 1:39150 to 250 words each, and contain
43. 1:42factual information. Avoid including
44. 1:44tutorials or deployment details. This is
45. 1:47an example of what a claude.md file
46. 1:49looks like in markdown format. It
47. 1:51clearly outlines the project's
48. 1:53architecture, key technologies, current
49. 1:55migration status, known issues, and code
50. 1:58conventions, all providing critical
51. 2:00context. To verify that Claude code is
52. 2:02reading your file correctly, run the
53. 2:04Claude/status command. Look for key
54. 2:07indicators like Claude.md showing over
55. 2:10500 tokens, code availability above 10K,
56. 2:13and token variants under 50. Remember to
57. 2:16update this file regularly after
58. 2:17refactors. Persistent context improves
59. 2:20team workflows significantly. Metrics
60. 2:23like clarity, stability, and consistency
61. 2:25are improved. The team contract ensures
62. 2:28the same baseline context, consistent
63. 2:30suggestions, faster onboarding, and
64. 2:32shared ownership of the project context.
65. 2:35You can implement Claude.md in about 30
66. 2:37minutes. The steps involved creating the
67. 2:39file, populating the three sections,
68. 2:42running Claude/status to check, testing
69. 2:45with a small task, and finally
70. 2:46committing and pushing your work. This
71. 2:48effort yields 20 to 40% token savings
72. 2:51per session. The key takeaways are:
73. 2:53Establish a contract by documenting
74. 2:55once, realize efficiency by saving over
75. 2:583,200 tokens per session, keep fresh by
76. 3:02updating every 2 weeks, and maintain
77. 3:04consistency so your team shares the same
78. 3:06baseline understanding across the three
79. 3:08sections: Overview, Tech, and Current
80. 3:11State. For resources, you can find the
81. 3:13docs at systemrd.com
82. 3:16and examples on GitHub. Run the
83. 3:19Claude/status tool to verify your setup.
84. 3:22Our next lesson is 2.1, Context
85. 3:25Stacking. We also have suggested
86. 3:27homework to deepen your understanding.